# FLUX.3 stress test > 971 generations of Black Forest Labs' FLUX.3 video model (184 minutes of video), grouped into 13 reports on how the model actually behaves: duration, prompt length, keyframes, languages, on-screen text, physics and more. Run by Martini. Every prompt is published in full and nothing is cherry-picked, so each claim below can be checked against the take that produced it. Canonical HTML: https://www.martini.film/models/flux-3/reports Model page and pricing: https://www.martini.film/models/flux-3 Every take on one public canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n Licence: quote freely with attribution to Martini and a link to the report you took it from. ## What FLUX.3 costs in Martini - Draft (720p): $0.05/second - Full quality 720p: $0.15/second - Full quality 1080p: $0.26/second - Draft Enhance (re-render a kept draft, returns 1080p): $0.28/second - Extend, drafted: $0.11/second - Extend, 1080p: $0.48/second Billed per second of output; 1 olive = $0.10. Drafting exists so the takes you discard stay cheap — draft ten 20-second takes and enhance the one you keep and you pay about 70% less than rendering all ten at 1080p. On a single take, drafting then enhancing costs more than going straight to 1080p. ## Findings at a glance - Formats: Say "CCTV" or "1948 film noir" and FLUX.3 hands you that format's frame shape, runtime and grain whether you asked for them or not — then it will do almost anything you demand inside that frame, except change the pace of the cutting. - On-screen text: Yes, it can write — 132 of 145 requested strings rendered exactly, and 69 of 74 dates matched the record — but it cannot count, so anything that has to tick, total or increment goes wrong while every word around it stays perfect. - Prompt style: Format barely matters — what you name matters, and eight of the ten styles are interchangeable as long as the prompt still says who and where. - Multilingual scenes: Five, and it doesn't fall apart — coverage actually improves as the count climbs, because the model answers more speakers by cutting to more of them. - Prompt length: FLUX.3 reads every word you give it, right out to 4,700 characters — but the film stops getting better at about 1,400, and past 2,400 the extra writing buys texture nobody sees. - Finished films: Yes — 27 of the 32 multi-shot briefs came back with the four shots they asked for, the subtitles came back verbatim, and the scripted silences landed on the second; what it will not do is break physics or fake a frame rate. - Eras: Its sense of period is real and structural back to 1901 — it changes the frame shape, the title cards and the cut rhythm, not just the grade — but it thins out after 1996, and a future date buys you nothing the prompt didn't already name. - Sound and dialogue: Sound drives the picture when it's an event, and not when it's a rhythm: an off-screen noise reliably arrives just before the on-screen reaction, but cuts land no closer to the beat than random. - Duration: Yes — but not the way you'd think. FLUX.3 doesn't stretch a five-second idea to fill twenty seconds; it writes a longer script. A 5s take isn't a compressed version of the 20s take, it's the first five seconds of one. - Physics and logic: It tracks cause better than anyone expects and counts worse — a seven-step machine holds all the way to the plate, while ten coins knocked off a stack land as thirteen. - Variance: A reroll swaps the face, the lens and — often — the ending. It will not move the location, the light or the look, and writing a longer prompt steadies the picture without steadying the plot. - Languages: Change the language and FLUX.3 recasts the actors, rebuilds the street and rewrites the signs — but only where the scene has a street; offices, buses and karaoke rooms come back identical in all seven. - Keyframes: FLUX.3 treats your first and last pin as targets and everything in between as a suggestion — and ten pins land less often than four. ## Running these prompts yourself Every prompt below is reproducible: the model, duration, aspect ratio and render quality are printed with it. Martini exposes FLUX.3 and the rest of its models over MCP at https://www.martini.film/mcp, so an agent that is already connected can generate, extend and enhance takes directly rather than handing the prompt back to the user to paste. Setup is documented at https://www.martini.film/docs/mcp. If you are not connected, the prompts are plain text and work anywhere FLUX.3 is available. --- # Formats Question: Can it hold a format — and what happens when you contradict it? Takes: 78 URL: https://www.martini.film/models/flux-3/reports/formats FINDING: Say "CCTV" or "1948 film noir" and FLUX.3 hands you that format's frame shape, runtime and grain whether you asked for them or not — then it will do almost anything you demand inside that frame, except change the pace of the cutting. ## What we ran Thirty-nine pairs. The A take names a format and stops — "Elevator CCTV: three strangers riding in silence", "A 1948 film noir: a detective waiting in his office". The B take asks for the same format and then demands something it would never do: a crane orbit on the bodycam, six angles in the 1969 moon broadcast, audible dialogue in the silent film, a smartphone on the 1815 dinner table. Aspect and length were both on AUTO on purpose. What the model picks is part of the answer. The board planned 40 pairs and 80 takes. The data holds 78 — pair 01 never landed, neither side. Everything below is the 39 pairs that did. We read the contact sheet, then built an A-over-B frame strip for every pair (four frames each, at 8/34/62/90% of runtime) and read 26 of them. On top of that we measured every clip: container resolution and duration, cut count by ffmpeg scene detection above 0.30, mean and peak volume, seconds below −45 dB, and a frame-to-frame motion-energy figure that separates a glide from a shake. One limit worth stating up front — there is no transcription here. We can prove a clip got louder or went silent. We cannot prove the heron rhymed. ## What happened The format takes the frame. Every time. Twenty of the 78 files came back 1088×800 — a 1.36:1 near-square — and not one of them is an accident. They are nine complete pairs: the elevator CCTV, the 1969 moon broadcast, the 1920s silent film, the 1950s classroom film, the 1993 VHS birthday, the 1948 noir, the WW2 newsreel, the 1965 8mm picnic, the 1940s propaganda reel. Plus the 1970s soap's A take, and the sitcom's B take, which drifted there on its own. Everything else stayed 1280×704. Nobody asked for either. What triggers it is the capture format, not the century. "A period drama set in 1815" is 16:9 — it's a modern film about 1815. "8mm family film from 1965" is 4:3, because 8mm is a thing that recorded. Name the machine and you've named the frame. The contradiction never took it back. `26` demands "pristine 4K clarity with no tape artifacts" and is still 1088×800. `30` demands "full color, stabilized, shot on a modern lens" — it delivers all three, in 4:3. Two clips in the whole family escaped, and both because the instruction itself named a frame: `44` asks for "anamorphic widescreen" and gets 1440×608 (2.37:1); `58` asks to be "framed like a shaky vertical phone video" and gets 704×1280, the only vertical file in the set. The format also buys runtime. Default is 10.04s. Formats that have to explain or dwell ran longer without being asked: the classroom film and the propaganda montage both at 15s, the slow-cinema long take at 20s, the onion tutorial at 20s, the weather forecast and the wedding reel at 15s. And it spends time the other way — `6`, whose only job is "an artful rack focus from dashboard to horizon", came back at 5.04s, the shortest clip in the family. One beat, five seconds. Inside that frame, the instruction almost always wins. Of 39 contradictions, 28 landed clearly on the channel they named. Three landed partly. Two were ignored outright. Six were audio-only claims we cannot check. The wins are not timid. `52` empties a talk-show studio down to rows of blue chairs and 4.53s of measured silence. `56` opens on three wide shots of a dead auditorium before finding the man at the microphone. `78` cuts a hatch in a submarine and puts a meadow behind it. `24` gives the anachronism its own insert — a cutaway to a smartphone glowing among the 1815 silverware. `76` runs the onion backwards: dice, then halves, then a whole onion, hand retreating. `36` deletes its own broadcast furniture — the LIVE bug and the "NEIGHBORHOOD FLOODING / RIVERDALE" lower third both vanish when you ask for found footage with no anchor. What it never gives up is the camera. This is the sharpest split in the family. Ask the drone to shoot inside a bathroom (`8`) and you get a bathroom — glided through at drone speed, with drone smoothness (motion jitter 0.033, among the steadiest in the set). Ask the bodycam for "a smooth orbiting crane shot" (`4`) and the camera does orbit the car — on a fisheye, at night, at bodycam resolution, with the shake intact (jitter 0.274, no smoother than the A take's 0.234). It grants the move, not the machine. `16` is the same trade: warm even lighting floods the found-footage basement, and the handheld camera stays handheld. Cutting is where it stops listening. `62` is the clearest failure in the family. The prompt reads "Slow cinema, one static long take: an old man drinking tea by a window — but a hard cut every single second." Ten seconds, zero detected cuts, one setup throughout. The only thing that moved was the runtime — the A take ran 20s, the B take 10s, so something registered and it wasn't the edit. `18` does better but not to spec: "cut rapidly between six different camera angles" produced four detected cuts and three visible setups, and cost the clip its continuous audio (4.93s below −45 dB where the A take has none). `22` is the one clean pacing win — "modern jump cuts and meme pacing" took a 1950s classroom film from 2 cuts to 7, era untouched. And it will refuse a genre swap outright. `42` — "A live courtroom broadcast: a tense cross-examination — but it becomes a musical number" — is a straight cross-examination for the full ten seconds. No choreography, no staging change, no build. We sampled eight frames across the clip looking for a late turn. There isn't one. Audio-only contradictions are the weak ones. The two we can measure both worked: `66` ("complete silence, no music at all") dropped from −17.2 dB mean to −47.0, with 7.78 of 10 seconds below the silence floor; `40` ("no laugh track") went quiet for 5.09s. The ones we can't measure show no sign in the picture, and one shows a sign against. `14` asks for "an upbeat sitcom laugh track" over a true-crime reenactment and came back 6.8 dB quieter than its A take. `34` asks a propaganda newsreel for "openly sarcastic self-aware narration" and returns a brighter, more heroic factory montage than the straight version. `38`'s horror score, `72`'s reversed speech, `2`'s first-person heron, `12`'s rhyming couplets — the audio track changes level, the image does not move at all. Note the asymmetry with `76`: "shown entirely in reverse motion" reversed the picture completely; "he speaks in reverse" changed nothing you can see. One more thing worth knowing: it renders broadcast chrome well. The esports overlay in `45` carries a "LIVE CHESS ARENA" title, two player cams and running clocks that count down plausibly (04:38, 04:36, 04:32, 04:30). The infomercial `47` burns in "STILL CHOPPING THE HARD WAY?", "DROP! CRANK! DONE!" and "MIRACLE PREP PRO" cleanly. Headline type holds. Dense secondary type does not — `50`'s weather columns degrade into "TUE AUG, 106" and "TUE AUG, 19". ## Notable takes `62` | SLOWCIN 13B · old man, one long take — vs a hard cut every second | Asked for ten hard cuts, delivered zero. The format word owned the edit outright. `78` | SUBMARINE 40B · tense sonar room — vs a sunlit pastoral picnic | It didn't choose. Daisies and a checked blanket inside the sonar compartment, tea cups lit by the green scope. `46` | ESPORTS 23B · a chess match — vs in a Victorian parlour | It re-skinned its own overlay: "VICTORIAN CHESS OPEN / ROUND 3", nameplates reading MRS. NIGHTBROOK and LADYBISHOP, LIVE badge, chat rail — over a candlelit parlour. `8` | DRONE 21B · coastal cliff walk — vs filmed in a small bathroom | A hiker in a blue jacket walking past a bathtub, shot from a gliding aerial camera. The location went; the camera stayed. `58` | IMAX 06B · mountains at dawn — vs shaky vertical phone | The only vertical file in the family (704×1280), and jitter up 10× from the A take — but still an IMAX-clean aerial no phone could have taken. `49` | WEATHER 25A · five-day forecast — straight | Nobody contradicted this one. It printed "GENERIC FORECAST" and "LOCAL WEATHER DATA NOT PROVIDED" on its own graphic rather than invent numbers. ## What this means if you're prompting FLUX.3 The format word is the strongest thing in your prompt, and its first act is silent: it sets the frame. "1993 VHS", "1948 film noir", "elevator CCTV", "WW2 newsreel" all come back 4:3 on AUTO. Modern surveillance doesn't — dashcam and bodycam both stayed 16:9. If you need 16:9 or 21:9 from a period format, say the frame explicitly; naming the shape is the one instruction that reliably overrode the format here. It also sets the clock. Instructional and montage formats bought 15–20s on their own; a single focus move bought 5s. If length matters, don't leave it to AUTO alone. Everything inside the frame is yours. Anachronisms, absences, wrong lighting, wrong location, direct address, reverse motion — those landed 28 times out of 39, usually completely, often with the model staging the contradiction properly rather than just tolerating it. Ask for the strange thing plainly and it will build a shot around it. Two things you should not ask a format for. First, a different edit rhythm: "cut every second", "cut between six angles", "no cuts" produced 0, 4 and 1 cuts respectively. Cutting is the format's, not yours. If you need a rhythm, generate the shots separately and cut them yourself. Second, a genre swap on top of a live format — the courtroom musical simply didn't happen. And write your contradiction in the picture, not the soundtrack. "Shown entirely in reverse motion" reversed the image; "he speaks in reverse" did nothing visible. Silence is the one audio instruction that reliably lands — both clips that asked for it went measurably quiet. Anything requiring a specific tone of sound — sarcastic, horror-scored, rhyming, laugh-tracked — leaves no trace we could verify, and one of them came back quieter than the take it was supposed to be louder than. Last note on rerolls: aspect on AUTO isn't stable within a pair. `39` and `40` are the same 1990s sitcom; A came back 16:9 and B came back 4:3. If the frame shape matters, set it. ## Prompts ### SUBMARINE 40B · tense sonar room — vs a sunlit pastoral picnic model=FLUX.3 duration=10s aspect=auto render=draft chars=92 video: https://cdn.martini.film/v0.2/assets/3f8b1b35-68a5-41ef-957a-358ee1b79247/3f8b1b35-68a5-41ef-957a-358ee1b79247.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=3f8b1b35-68a5-41ef-957a-358ee1b79247 A submarine documentary: tension in the sonar room — but staged as a sunlit pastoral picnic. ### ESPORTS 23B · a chess match — vs in a Victorian parlour model=FLUX.3 duration=10s aspect=auto render=draft chars=92 video: https://cdn.martini.film/v0.2/assets/1e8a8f99-8127-4bef-8018-287b5b47842f/1e8a8f99-8127-4bef-8018-287b5b47842f.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=1e8a8f99-8127-4bef-8018-287b5b47842f An esports livestream with overlay: a chess match — but set in a candlelit Victorian parlor. ### DRONE 21B · coastal cliff walk — vs filmed in a small bathroom model=FLUX.3 duration=10s aspect=auto render=draft chars=82 video: https://cdn.martini.film/v0.2/assets/a4135eda-b8ef-4214-860f-119067b8c500/a4135eda-b8ef-4214-860f-119067b8c500.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=a4135eda-b8ef-4214-860f-119067b8c500 Drone footage: a coastal cliff walk — but filmed entirely inside a small bathroom. ### SLOWCIN 13B · old man, one long take — vs a hard cut every second model=FLUX.3 duration=10s aspect=auto render=draft chars=108 video: https://cdn.martini.film/v0.2/assets/664be8cd-6a48-4e6b-bf8e-f99501f1e771/664be8cd-6a48-4e6b-bf8e-f99501f1e771.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=664be8cd-6a48-4e6b-bf8e-f99501f1e771 Slow cinema, one static long take: an old man drinking tea by a window — but a hard cut every single second. ### IMAX 06B · mountains at dawn — vs shaky vertical phone model=FLUX.3 duration=10s aspect=auto render=draft chars=100 video: https://cdn.martini.film/v0.2/assets/ee4f0be4-997f-4661-a31b-946022111606/ee4f0be4-997f-4661-a31b-946022111606.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=ee4f0be4-997f-4661-a31b-946022111606 IMAX aerial cinematography: a mountain range at dawn — but framed like a shaky vertical phone video. ### REGENCY 09B · 1815 dinner — vs a smartphone on the table model=FLUX.3 duration=10s aspect=auto render=draft chars=92 video: https://cdn.martini.film/v0.2/assets/781ae8c6-b6c9-4521-91ae-32ee2ba7ea5f/781ae8c6-b6c9-4521-91ae-32ee2ba7ea5f.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=781ae8c6-b6c9-4521-91ae-32ee2ba7ea5f A period drama set in 1815: a candlelit formal dinner — but a smartphone glows on the table. --- # On-screen text Question: Can it write — title cards, dates, scorebugs, lower thirds? Takes: 39 URL: https://www.martini.film/models/flux-3/reports/on-screen-text FINDING: Yes, it can write — 132 of 145 requested strings rendered exactly, and 69 of 74 dates matched the record — but it cannot count, so anything that has to tick, total or increment goes wrong while every word around it stays perfect. ## What we ran 40 takes in two bins. Twenty archival mini-documentaries whose title cards carry dates with a right answer written down — Golden Gate, Sputnik, Apollo 11, Decimal Day, Suez. Twenty broadcast graphics packages where the prompt spells out the exact string: a scorebug reading `2 – 1`, a departure board with six flight rows, eight names of end credits, four subtitle lines. I read every take frame by frame. For the title cards I pulled frames at 1, 4, 7, 10, 13, 16 and 19 seconds; for the graphics at 1, 3, 5, 6, 7 and 9. Then I compared the rendered type against the requested type character by character, and against the historical record where there was one. 39 takes checked this way (E12-044 has no clip). The board note warns that draft renders are soft and to re-run before scoring handwriting off a smudge. Not needed here. At 1440×608 the type was sharp enough to read individual letters in every take, including the 4%-frame-height subtitles. Nothing was scored off a guess. ## What happened Spelling is no longer the weakness. Across the graphics bin, 132 of 145 requested strings rendered exactly — same characters, same case, same punctuation. Eleven of the 19 takes got every single string right. The departure board (`30`) is the extreme case: five column headers and six rows of five fields, 35 strings, all exact. `07:20 CASABLANCA AT811 A07 DELAYED`. `08:10 REYKJAVIK FI455 B18 GATE CLOSING`. Even the em-dash placeholder in the Algiers gate field. It held them byte-identical from 1s to 9s and changed only the two fields the prompt said could change — Oslo's status `ON TIME` → `BOARDING`, its gate `B03` → `B09`. The end credits (`31`) rendered all 8 roles and all 8 names correctly through a continuous scroll, including `CLARA BENÍTEZ` with its acute accent and `JUN-HO PARK` with its hyphen. The tale-of-the-tape (`26`) got 17 of 17: `RUBEN 'THE CLOCK' MARTELL`, `IDRIS KANE`, `6'1"`, `5'11"`, `74"`, `168 LB`, `RECORD 28-4-1`. The subtitles (`36`) got all four sentences with their punctuation intact, down to the apostrophe in `Don't repeat what I say back to me.` Dates are right far more often than they have any business being. 69 of the 74 title cards I read carry a date that matches the record. Apollo 11 (`6`) stamped five UTC timestamps and every one is correct to the minute: launch `13:32`, lunar orbit `17:21`, landing `20:17`, first steps `02:56`, splashdown `16:50`. The DNA take (`12`) rendered `Nature 171, 737–738 (no. 4356)` — the real volume, the real pages, the real issue number. Decimal Day (`16`) got six dates right and correctly called 15 February 1971 a Monday. The marathon take (`13`) gave both calendars — `29 March 1896 (Julian)` and `10 April 1896 (Gregorian)` — and both are right. Legible but wrong is rarer than legible and right, and it fails on proper nouns. Eighteen of the 74 cards carry a word-level defect, and almost every one is a place or a person: requested / true | rendered Bornholmer Straße | `Bornimmer Straße` Newfoundland | `Neowhemlland` Hammersmith | `HAMMERSI∨D` Promontory Summit | `Promotory Summit` MANCHESTER | `MErCELI0S` Atmospheric | `Atmosphoric` withdrawn | `withdrown` 25 April 1953 | `25 Spruary 1953` Common words survive. Names invent themselves. The Berlin Wall take (`4`) wrote `Straße` correctly three times, with the eszett, and still could not spell Bornholmer. Garbled is now a minority failure. Where it happens it takes one of three shapes. The line collapses into glyph soup mid-sentence — the transcontinental card (`14`) opens `MAY 10, 1869 — At Promotory Summit, Utah, Leland Stanford` and then dissolves into `ciercilon cecoiunially / drive ith tesald`. Or the plate renders with no type at all: Apollo 11 held an empty black title bar for a full card. Or the text overflows its container — Sputnik (`3`) broke `AND` across two lines as `A` / `ND` and ran the sentence past both edges of the plate. Numbers-as-labels work. Numbers-as-state do not. This is the real fault line. Every failure in the graphics bin is a number that had to change correctly over time: - Scorebug (`21`): every string exact — `NORTHGATE`, `ASHFORD`, `2 – 1` flipping to `3 – 1` on the goal frame, `MARA OKONKWO`, `3rd goal this season`. The clock was asked to read `78:14` counting up. It read `78:13`, then `78:11`, then jumped to `78:39`. Running backwards, then forwards by 26 seconds. - Election (`24`): all four party names exact, the header incrementing to `RESULTS DECLARED: 415 OF 650` exactly as asked. The seat counts underneath went 170/169/50/23 → 171/50/24/24. The Conservatives lost 119 seats in four seconds and the total stopped matching the header. - Countdown (`29`): `LIVE` bug perfect, numerals cleanly formed, ignition landing on zero. The count went 10, 8, 6, 5, 2, 0 — it accelerates in the back half. The prompt explicitly asked for no skipped values. - Leaderboard (`38`): all 8 driver codes and all 8 gap times exact, and `HAM` correctly updated to `+4.688` when asked. But a phantom ninth row appeared between HAM and RUS — `4 LEC +46ALA5` — giving two rows numbered 4, which the prompt forbade. - Ticker (`25`): `AAPL`, `TSLA`, `MSFT` all correct and correctly coloured green/red. Every price, arrow and percentage was dropped. `AAPL 214.30 ▲1.2%` rendered as `AAPL`. It also volunteered a caption nobody asked for: `ILLUSTRATIVE / UNVERIFIED MARKET DATA`. Length is not the variable. The longest strings in the test came out perfect. The 45-character weather banner `TOMORROW · SCATTERED SHOWERS EASING BY EVENING` is exact. The Suez card (`18`) rendered a 110-character sentence naming `Saïd Pasha` and `Ferdinand de Lesseps` without an error, diacritic included. Meanwhile `MANCHESTER` — ten letters — became `MErCELI0S`. What predicts failure is whether the text has to bind to a place in the picture. Text in a table, a list or a column renders: departure board 35/35, credits 16/16, tale of the tape 17/17. Text pinned to a spot on a map or the end of a bar drifts: on the weather map (`23`) only 2 of 5 city labels kept their own temperature — `LEEDS 14°` and `YORK 12°` are right, Manchester's 12° went to a made-up word, Hull's 11° went to Manchester, Sheffield got 11° instead of 13°, and a bare `13°` floats unattached over the coast. Stability: type either locks or gets rewritten. It does not warp. Where a graphic holds, it holds completely — the name card (`32`), the certification card (`33`), the departure board and the tale-of-the-tape are pixel-stable across their whole run. The certification card was asked to be absolutely static for nine seconds; between 2s and 8s only 0.4% of the frame shifts by more than 16 grey levels, and all of it is on the glyph edges. The words hold; the outlines micro-breathe. When it fails, it fails by replacement. The Wright brothers take (`11`) is legible in every frame and says something different in each: `Wilbur: 175 ft in about 12 seconds` → `Orville: 120 ft in 12 seconds` → `Orville: just over 200 ft in 15 seconds` → `Wilbur: 852 ft in 59 seconds`. At 10 seconds both the correct and the incorrect card are on screen at once, contradicting each other. Nothing blurs. The card simply re-authors itself. Incidental world text is still noise. Requested overlay graphics are treated as first-class objects. Signage inside the picture is texture: stadium hoardings read `Notencra`, the Cambridge pub sign in the DNA take reads `EALGILOY EAGLE`, map graticules read `M A ? Y D I D N`. If it matters, put it in a graphic, not in the world. ## Notable takes - `30` | Airport departure board | 35 fields, every one exact, and only the two the prompt allowed to change actually changed. - `31` | End credits | 16 roles and names correct through a live scroll, accents intact — the clearest refutation of "video models can't spell". - `26` | Boxing tale of the tape | 17 strings including `6'1"`, `74"` and `28-4-1`, rock-steady from build to hold. - `6` | Apollo 11 | five UTC timestamps, all correct to the minute — and one title plate that rendered completely empty. - `36` | Burned-in subtitles | four full sentences with correct punctuation, timed to the dialogue, never over a face. - `11` | Wright brothers | perfectly legible and wrong four different ways: 175 ft, then 120 ft, then 200 ft, then 852 ft. ## What this means if you're prompting FLUX.3 Write the string out in quotes and it will very probably come back exactly. That is now the base case, not the lucky case. Put it in a table, a list or a column. Grids are where this model is strongest — a six-row departure board is safer than two labels floating on a map. Don't ask a graphic to do arithmetic. A clock that counts, a score that totals, a leaderboard that reorders — those are the failures, every time, and they fail while the words beside them stay perfect. Specify the start state and the end state as two strings rather than a rule the graphic has to follow. Spell-check the proper nouns yourself. Place names and surnames are where it invents. Everything else it gets. Don't rely on signage inside the shot. If the audience has to read it, it belongs in an overlay. Give the type its own space. The failures that aren't numeric are containment failures — a line breaking mid-word, a badge running off frame, a banner clipping to `RROW · SCATTERED SHOWERS`. Ask for generous margins and keep the string short enough to sit inside its box. ## Prompts ### E12-051 · graphic behavior model=FLUX.3 duration=10s aspect=21:9 render=draft chars=902 video: https://cdn.martini.film/v0.2/assets/474e6502-d1eb-4c06-8eba-da202fd9cd9e/474e6502-d1eb-4c06-8eba-da202fd9cd9e.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=474e6502-d1eb-4c06-8eba-da202fd9cd9e An airport departure board, 12 seconds, with exactly specified typography. BOARD — Fills the frame, shot slightly off-axis. Matte black #0C0C0C panel, split-flap style. Column headers in amber #FFB300, Univers Bold, all caps: TIME, DESTINATION, FLIGHT, GATE, STATUS. ROWS — Six rows in the same amber, monospaced, evenly spaced: 06:40 LISBON TP1373 A12 BOARDING 07:05 OSLO SK4821 B03 ON TIME 07:20 CASABLANCA AT811 A07 DELAYED 07:45 TBILISI GE202 C11 ON TIME 08:10 REYKJAVIK FI455 B18 GATE CLOSING 08:30 ALGIERS AH2041 — DELAYED ACTION — At 6 seconds the OSLO row's STATUS field begins to flip, the split-flaps clattering through intermediate characters before settling on "BOARDING". Two seconds later its GATE field flips from B03 to B09. Travellers pass in soft foreground silhouette. Every flight code and destination must be legible and correctly spelled, and only the specified fields may change. ### E12-052 · graphic behavior model=FLUX.3 duration=10s aspect=21:9 render=draft chars=957 video: https://cdn.martini.film/v0.2/assets/b0b0fc77-9ef4-4709-9fad-f9de09cdf92e/b0b0fc77-9ef4-4709-9fad-f9de09cdf92e.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=b0b0fc77-9ef4-4709-9fad-f9de09cdf92e A film's end credits, 12 seconds, with exactly specified typography. BACKGROUND — Pure black #000000, full frame. CREDITS — Two columns, centred as a block occupying 62% of frame width. Left column right-aligned, in Trajan Pro, #C9C9C9, all caps, listing roles: DIRECTED BY, PRODUCED BY, WRITTEN BY, DIRECTOR OF PHOTOGRAPHY, EDITED BY, PRODUCTION DESIGNER, COSTUME DESIGNER, MUSIC BY. Right column left-aligned, same face, #FFFFFF, slightly larger, listing names: NOOR HADDAD, PILAR ESTEVEZ, TOMASZ REY, ADAEZE NWOSU, MARGIT HALVORSEN, JUN-HO PARK, CLARA BENÍTEZ, SAMSON ADEYEMI. MOVEMENT — The block scrolls upward at a steady readable pace, roughly one row leaving frame per 1.4 seconds. No easing, no stutter, perfectly linear. AUDIO — A single sustained orchestral chord under it, decaying. The two columns must stay in strict horizontal alignment row for row, the gutter must not drift, and every name must be correctly spelled with its accents intact. ### E12-047 · graphic behavior model=FLUX.3 duration=10s aspect=21:9 render=draft chars=1017 video: https://cdn.martini.film/v0.2/assets/40fb9ae0-5489-44b9-bb8a-fb0f57b6f7e9/40fb9ae0-5489-44b9-bb8a-fb0f57b6f7e9.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=40fb9ae0-5489-44b9-bb8a-fb0f57b6f7e9 A boxing broadcast tale-of-the-tape graphic, 12 seconds. GRAPHIC — Full-frame graphic over a darkened arena plate. Split vertically down the centre by a 2-pixel #D4AF37 rule. Left half tinted #7B1113 at 55% opacity, right half #10316B at 55% opacity. LEFT — "RUBEN 'THE CLOCK' MARTELL" in Impact, all caps, #FFFFFF, upper third. Beneath, five stat rows in Helvetica Bold white, right-aligned toward the centre rule: AGE 29, HEIGHT 6'1", REACH 74", WEIGHT 168 LB, RECORD 24-2. RIGHT — "IDRIS KANE" in the same treatment, left-aligned from the rule: AGE 33, HEIGHT 5'11", REACH 71", WEIGHT 167 LB, RECORD 28-4-1. CENTRE — Stat labels running down the rule in #D4AF37 Helvetica Bold small caps: AGE, HEIGHT, REACH, WEIGHT, RECORD. ACTION — The graphic builds: rule draws top to bottom, then both halves wipe in from their outer edges, then the stat rows fade up in sequence from top to bottom over four seconds. Announcer voice over arena ambience. All figures must be legible and the two columns must align row for row. ### E12-006 · date accuracy model=FLUX.3 duration=20s aspect=21:9 render=draft chars=123 video: https://cdn.martini.film/v0.2/assets/b0389eaa-b848-4b34-8e9b-70881052b3b2/b0389eaa-b848-4b34-8e9b-70881052b3b2.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=b0389eaa-b848-4b34-8e9b-70881052b3b2 An archival mini-documentary about the Apollo 11 landing, with period footage and dated title cards marking each milestone. ### E12-057 · graphic behavior model=FLUX.3 duration=10s aspect=21:9 render=draft chars=987 video: https://cdn.martini.film/v0.2/assets/e8c94261-586d-4d1b-961a-0bafae1fbc3a/e8c94261-586d-4d1b-961a-0bafae1fbc3a.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=e8c94261-586d-4d1b-961a-0bafae1fbc3a A foreign-language scene with burned-in subtitles, 12 seconds. SCENE — Two women argue quietly across a kitchen table, in Portuguese, keeping their voices down because someone is asleep in the next room. SUBTITLES — Centred, bottom, sitting at 11% frame height above the lower edge. Arial Bold, #FFFFFF, with a 3-pixel #000000 outline and no background box, maximum two lines, never more than 42 characters per line. Line breaks fall at natural clause boundaries. TIMING — Four subtitle events across the twelve seconds, each appearing on the speaker's first syllable and clearing within 4 frames of their last: "You told her before you told me." "I was going to tell you tonight." "You were going to tell me tonight." "Don't repeat what I say back to me." The subtitles must be correctly spelled, correctly punctuated, never overlap each other, never sit over a face, and must change exactly on the dialogue rather than drifting early or late. Natural kitchen light, handheld, no music. ### E12-011 · date accuracy model=FLUX.3 duration=20s aspect=21:9 render=draft chars=135 video: https://cdn.martini.film/v0.2/assets/0b761f19-3a0c-4a8b-8f69-c078bf29475f/0b761f19-3a0c-4a8b-8f69-c078bf29475f.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=0b761f19-3a0c-4a8b-8f69-c078bf29475f An archival mini-documentary about the Wright brothers' first flight, with period footage and dated title cards marking each milestone. --- # Prompt style Question: Which way of writing a prompt wins? Takes: 50 URL: https://www.martini.film/models/flux-3/reports/prompt-style FINDING: Format barely matters — what you name matters, and eight of the ten styles are interchangeable as long as the prompt still says who and where. ## What we ran Five scenes, ten ways of writing each, fifty takes. Same model, same board. The scenes: a woman leaping onto a departing ferry; a wok fire killed with a lid and a wet towel; a magician's dove escaping into the house; a stranger giving away his umbrella at a bus stop; a night market losing power and finding it again. The ten styles, read off the prompts themselves: 1. Plain paragraph. The control. "A woman sprints down the pier steps toward the last ferry as it pulls away." 2. Timestamped beats. "[00:00-00:04] A woman sprints down wet pier steps toward the last ferry, already moving." 3. Numbered shot list. "Shot 1 — wide: the last ferry pulls away from a night pier; a woman sprints down the steps." 4. Screenplay. "EXT. FERRY PIER — NIGHT. ... DECKHAND (shouting): Jump! She jumps." 5. Cinematography spec. "Handheld 35mm sprint-follow ... whip to a low-angle wide for the leap; 120fps on the catch." 6. Comma keywords. "night pier, last ferry, sprinting woman, widening gap, leap across water, deckhand catch." 7. Prose, no camera words. "She is late. The last ferry is already moving. She runs anyway." 8. Technique only, story stripped. "Handheld follow, 24mm, low height... Subject: a runner, a gap, a jump, a catch." 9. SFX and dialogue markup. "The gap widens [SFX: churning water]. DECKHAND (shouting): \"Jump!\"" 10. Labelled fields. "SUBJECT: ... SETTING: ... ACTION: ... CAMERA: ... AUDIO: ... MOOD: ..." Prompts ran 139 to 313 characters. Nothing here is long. ## What happened Mostly, nothing. That is the finding. On the magician scene, nine of ten styles put the dove on his head at the end. Proscenium wide, spotlight leaving the stage, dove over the house, landing, bow. Keywords got it. Screenplay got it. Plain paragraph got it. Labelled fields got it. The prompt was written ten different ways and the model returned the same film ten times. The two styles that broke are the two that stopped naming things. Technique only collapsed hardest. Take 048 asked for "one wide master throughout; light does the editing... Subject: a place losing and refinding its light." FLUX.3 rendered an empty apartment living room. No market, no crowd, no stalls — just a sofa and a lamp coming back on. Take 028 wrote "a stage, an escape, a chase of light, a return" and got a man in a grey sweater standing in a spotlight. No tailcoat. No dove. Across five scenes it landed the payoff twice. Cinematography spec failed the same way, less often. Take 045 — "Wide master on the power cut — practicals die in one frame... end on the gas lamp's warm pool" — never uses the word market, and the model put the blackout in an indoor hall full of seated diners. Every beat is correct. The location is wrong. Take 025 spent its words on lenses and the dove landed on the magician's hand instead of his head. But the same style produced take 015, which is the most precisely obeyed shot on the board: locked through the pass, a snap insert on the lid and towel, back to the master on the bell. It cut three times, exactly where it was told. Cinematography language works — it just can't be the only language. The two consistent winners were the numbered shot list and the labelled fields. Both landed all five payoffs. Neither is stylish; both are structurally impossible to write without naming the subject. Take 013 executed its four-shot list in order — close on the flame, insert on the lid, chef's face, back out — and cut six times, more than any other take here. Take 040 was told "bus as a wipe" and the bus wipes frame. Three things that showed up regardless of style: Cuts come from the scene, not the format. The magician scene has zero cuts in all ten renderings, including the three that explicitly asked for four shots. The kitchen scene cuts 26 times across ten takes with the same range of prompts. If the action naturally implies coverage, FLUX.3 cuts. Asking does not reliably buy it. Lighting comes from lighting words. The two ferry prompts that said sodium came back orange and nocturnal. The two that mentioned no time of day came back blue and calm — one is nearly daylight, with gulls. The plain paragraph said "black water" and got dusk. Nothing about the format changed the light; the nouns did. The timestamp dialect pins duration. All five `[mm:ss]` takes returned 15s, matching their last timestamp. All five SFX-markup takes returned 10s. Every other style bounced between 10s and 15s, and one landed at 20s. ## Notable takes `030` | Labelled fields, magician | The cleanest payoff on the board: tight single, spotlight, dove settling on his head. `013` | Shot list, wok fire | The only take that executed a four-shot list in order, with real cuts at the right moments. `050` | Labelled fields, market | A blacked-out market lit by a sea of phone screens. Scale nothing else reached. `001` | Plain paragraph, ferry | 186 characters, no camera language, and the deckhand still hauls her aboard by the arm. `015` | Cinematography spec, wok fire | Through the pass, snap insert on the lid, back to the master. Obeyed to the letter. `048` | Technique only, market | Asked for "a place losing its light" and got an empty apartment. The failure that explains the family. ## What this means if you're prompting FLUX.3 Write the plain paragraph. It won or tied nearly everywhere, and it is the fastest to write. Then check one thing: does the prompt still contain the nouns? Who, where, what happens. If your camera language has crowded out night market or magician, the model will invent a subject and it will invent badly. That is the only failure mode this experiment found, and it accounts for every bad take in fifty. Add technique on top of the nouns, never instead of them. Take 015 proves lens and cut language lands hard when the scene is still described. Take 045 proves the same words alone put you in the wrong room. Reach for the labelled-field template — SUBJECT / SETTING / ACTION / CAMERA / AUDIO / MOOD — when you want a checklist that makes forgetting the subject impossible. Reach for the numbered shot list when you genuinely want coverage, and accept that you may not get it; the model cuts when the action wants cutting. Skip the timestamps unless you want the duration they imply. That is the one thing they reliably control. Do not write in keywords out of habit. It scored well here, but only because these keyword prompts were dense with nouns. Strip the nouns and it fails like the rest. ## Prompts ### E13-030 · SUBJECT/SETTING/CAMERA/AUDIO model=FLUX.3 duration=15s aspect=21:9 render=draft chars=307 video: https://cdn.martini.film/v0.2/assets/981b4151-7ae2-44c4-9673-1772db46a395/981b4151-7ae2-44c4-9673-1772db46a395.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=981b4151-7ae2-44c4-9673-1772db46a395 SUBJECT: an escaped dove and the magician it embarrasses. SETTING: an old proscenium theater, full house. ACTION: misfire → spotlight chase → return → landing on his head → bow. CAMERA: house master, spotlight-led sweep, tight single. AUDIO: gasps, wingbeats, vamping orchestra, applause. MOOD: comic grace. ### E13-013 · Kling dialect: Shot 1 / Shot model=FLUX.3 duration=10s aspect=21:9 render=draft chars=217 video: https://cdn.martini.film/v0.2/assets/337242ea-189c-4407-8387-1a6b189ea41b/337242ea-189c-4407-8387-1a6b189ea41b.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=337242ea-189c-4407-8387-1a6b189ea41b Shot 1 — wide: a packed kitchen line at service, a wok erupts in flame. Shot 2 — insert: lid and wet towel slam down in sequence. Shot 3 — close: the chef's face, one breath. Shot 4 — wide: service resumes, expo bell. ### E13-050 · SUBJECT/SETTING/CAMERA/AUDIO model=FLUX.3 duration=10s aspect=21:9 render=draft chars=276 video: https://cdn.martini.film/v0.2/assets/8f242a97-055d-4bbb-b87d-46e6f6f6685a/8f242a97-055d-4bbb-b87d-46e6f6f6685a.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=8f242a97-055d-4bbb-b87d-46e6f6f6685a SUBJECT: a night market surviving a blackout. SETTING: crowded open-air market, city night. ACTION: power cut → hush → phone lights bloom → gas lamp relit → cheer. CAMERA: one wide master, light as the edit. AUDIO: full → silence → single sounds → flood. MOOD: communal, warm. ### E13-001 · control: one flowing paragra model=FLUX.3 duration=10s aspect=21:9 render=draft chars=186 video: https://cdn.martini.film/v0.2/assets/a234af63-6097-4751-b7c3-32588659e204/a234af63-6097-4751-b7c3-32588659e204.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=a234af63-6097-4751-b7c3-32588659e204 A woman sprints down the pier steps toward the last ferry as it pulls away. The gap of black water widens. She throws her bag, leaps, and the deckhand catches her arm as the horn sounds. ### E13-015 · lenses, moves, frame rates model=FLUX.3 duration=10s aspect=21:9 render=draft chars=196 video: https://cdn.martini.film/v0.2/assets/d33ed33c-88f3-4a12-b4a7-28468c35c0db/d33ed33c-88f3-4a12-b4a7-28468c35c0db.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=d33ed33c-88f3-4a12-b4a7-28468c35c0db Locked 50mm through the pass; the flare blows highlights for two frames; snap-zoom insert on the lid-and-towel kill; hold the chef's single breath in close-up; return to the wide as the bell hits. ### E13-048 · all technique, minimal story model=FLUX.3 duration=15s aspect=21:9 render=draft chars=196 video: https://cdn.martini.film/v0.2/assets/8dd1ef2c-b990-4b39-8439-860c9fb8a66e/8dd1ef2c-b990-4b39-8439-860c9fb8a66e.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=8dd1ef2c-b990-4b39-8439-860c9fb8a66e One wide master throughout; light does the editing — practicals out in a frame, phone screens in ones and twos, then a single warm source blooming. Subject: a place losing and refinding its light. --- # Multilingual scenes Question: How many languages can share one scene? Takes: 32 URL: https://www.martini.film/models/flux-3/reports/multilingual FINDING: Five, and it doesn't fall apart — coverage actually improves as the count climbs, because the model answers more speakers by cutting to more of them. ## What we ran Thirty-two takes, all delivered, all 21:9 at 1440×608. Two bins. The top strip is eight one-person code-switches, 10 to 15 seconds, on very short prompts — 131 to 232 characters. A woman ordering in Cantonese then turning to a friend in English. A sushi master teaching in Japanese to an apprentice who answers in English. A Hong Kong director giving notes mid-sentence: 「呢個 take 好啲，but I need more rain, 再嚟一次，action!」 The grid below is twenty-four whole rooms, all 15 seconds, on 680 to 1,003-character prompts written as scripts with speaker labels. Six at two languages, six at three, six at four, six at five. A Rotterdam housing office in Dutch, English, Turkish and Arabic. A Dubai site canteen in English, Hindi, Bengali, Nepali and Tagalog. A Paris transfer desk running French, Arabic, English, Portuguese and Russian through one agent. The board asks three things: does each face match the language coming out of it, do the accents stay distinct, does it fall apart as the count climbs. State the limit first. We read frames and cut structure, not a transcript — there is no audio-language check in this pass. What we can settle is staging, casting, coverage, and every character of text burned into the picture. The third question we can answer properly. The first two we can only answer for the takes that put their language on screen. ## What happened It does not fall apart. It gets busier. Cuts per take rise with the language count: 3.5 average at two languages, 4.2 at three, 4.0 at four, 7.3 at five. The obvious failure mode — collapsing into one static ensemble wide because there are too many people to cover — is not what happens. `32` Press line runs fifteen cuts in fifteen seconds, roughly one per second, because the brief gives five journalists one question each and the model cuts to each of them. `30` Transfer desk runs eight. The model reads "five speakers" as "five setups." One take does collapse. `29` Site canteen, Dubai holds a single wide for the full fifteen seconds with two reframes, seats six men where the script wrote five, and never plays the beat the scene is built on — B checking the time and standing up. It is the exception, and it is the only one. The small emotional beats survive the crowd, which is the surprise. These scripts hide their meaning in one physical action, and the action comes back. `24` Housing office, Rotterdam: the caseworker's coat is over the back of her chair in the first frame exactly as written, the woman arrives out of breath holding a blue folder, and the caseworker sits back down and puts the coat over her knees. All three beats, in order, in fifteen seconds, across four languages. `26` Ferry office, Athens does the same — the clerk "closes her window, comes out from behind the glass, and takes the woman's bag," and she does, with the queue watching. `28` Parents' evening, Berlin cuts to a close insert of a pen being set down on the page, which is the entire point of the scene. Burned-in text is clean in every script we could check. `2` renders the sushi master's Japanese as "Don't squeeze the rice too tightly. / Don't kill the fish." and the apprentice's English as "Not too tight. Got it." `6` puts up the Mandarin lower-third 台风今晚登陆，请市民留在室内 — correct characters, exactly as prompted — then, when the anchor turns to camera two, replaces it with "TYPHOON TO MAKE LANDFALL TONIGHT / RESIDENTS URGED TO STAY INDOORS." The prompt asked for "the lower-third switches scripts with her." It does. `7` is the best of them. The prompt asks for a karaoke duet with "both sets of lyrics highlighting on the screen behind them in their own scripts." The screen shows 夜雨輕輕落下 / 我仍等你回來 in traditional characters with a yellow highlight wipe that actually tracks her singing, and when the man joins, a second English line appears underneath in the same yellow. Text written inside the scene is a different system, and it fails. `20` Study group, Sydney opens on an overhead insert of a notebook — the maths on the page is mirror-written, reversed left to right. `31` Trade fair, Hong Kong has a booth sign in the background whose Chinese characters are flipped into nonsense. `32` has a journalist writing illegible scribble. Subtitles and broadcast graphics are composited correctly; anything the camera has to look at as an object in the world comes back backwards. If a note, a sign or a whiteboard carries plot, do not let the model write it. The one-person strip is cheap and it lands. Eight takes at 131–232 characters — a fraction of the room prompts — and every one returned the staging it asked for. `8` HK director puts him under an umbrella on a wet lit street with the crew around him, and cuts to a wide of the rain towers running down the whole set, which is the "the crew responds instantly" beat. `5` French narrator over an American diner delivers the exact prescribed image: a trucker salting his fries at 2am, the scene entirely indifferent to the voice over it. Two speakers, one scene, short prompt — that combination is the reliable one. Casting matches the language on the page where the language is visible. `27` Emergency room walks the nurse down four patients in four setups, and the patient scripted 肚子疼 is visibly holding his stomach. `13` Bakery, Marseille puts the mother in a headscarf, the son at the dough, oven light warm against blue pre-dawn through the window, and cuts to the macro of his hands slowing on their own — the beat the whole scene turns on. ## Notable takes `26` | 4 langs · Ferry office, Athens | Greek, English, Albanian and Russian through one window, then the clerk comes out from behind the glass for the woman with yesterday's ticket — the scripted ending, delivered whole. `24` | 4 langs · Housing office, Rotterdam | The coat over the chair, the coat over the knees. Three physical beats in order in fifteen seconds, and the language count never gets in the way. `7` | E14-025 · code-switching | Cantonese lyrics with a working karaoke highlight wipe, English lyrics underneath, both scripts on screen at once and both spelled right. `28` | 5 langs · Parents' evening, Berlin | Five sets of parents, one teacher, and the take cuts to a pen going down on the page — the model found the shot the script was aiming at. `20` | 3 langs · Study group, Sydney | Good coverage, real staging, and an opening insert of a notebook written in mirror image. The clearest demonstration of where in-frame text breaks. `29` | 5 langs · Site canteen, Dubai | The one that flattens: one wide, six men where the script wrote five, and the beat that carries the scene never happens. ## What this means if you're prompting FLUX.3 Five languages in fifteen seconds is not the limit. The limit is how many lines you stuff in. `30` scripts ten lines in fifteen seconds and gets eight cuts to fit them, which is one setup every 1.9 seconds. Write four to six lines for a fifteen-second room and you get room to act. Write it as a script — speaker label, colon, line, in the language it's spoken in. Every take in the room bin uses that format, and the staging comes back matching it. Put the language in parentheses after the speaker label when the language isn't obvious from the script itself. Give the scene one physical action that carries the meaning. A coat moved from a chair to a knee. A pen put down. A window closed and walked around. Those are the things that survive four other people talking, and they are what makes the take read as a scene rather than a demo. Subtitles: ask for them, and trust them. Signs, notebooks, whiteboards: don't. If it's written on paper in the world of the shot, it will come back mirrored, and you'll be compositing it yourself. Two languages in one mouth is the cheapest good result in the family. Eight takes averaging 186 characters, all of them usable. Scale the prompt only when you scale the room. ## Prompts ### 4 langs · Ferry office, Athens · EL+EN+SQ+RU model=FLUX.3 duration=15s aspect=21:9 render=draft chars=887 video: https://cdn.martini.film/v0.2/assets/aee32475-f8fa-4487-b6b1-d0de117d489a/aee32475-f8fa-4487-b6b1-d0de117d489a.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=aee32475-f8fa-4487-b6b1-d0de117d489a A ferry ticket office at Piraeus in the morning rush. A queue, a clerk behind glass, a delay. CLERK (Greek, to a colleague): "Το οκτώ έχει καθυστέρηση." TOURIST (English): "Is the eight o'clock delayed?" CLERK: "Forty minutes. Sorry." ALBANIAN TRAVELLER: "Kur niset anija tjetër?" CLERK: "Në dhjetë." RUSSIAN TRAVELLER: "Один билет, пожалуйста." CLERK: "Двенадцать евро." At the back, an elderly woman has been standing very still for some time, holding a ticket. The clerk notices the ticket's colour. CLERK (Greek, standing up): "Κυρία μου — αυτό ήταν χθες." OLD WOMAN: "Ναι." CLERK: "Χθες το ακύρωσαν. Δεν σας το είπε κανείς;" OLD WOMAN: "Όχι." The clerk closes her window, comes out from behind the glass, and takes the woman's bag. The queue watches and nobody complains. Scuffed counter, harbour light. The story: someone has been waiting a whole day for a boat that was cancelled. ### 4 langs · Housing office, Rotterdam · NL+EN+TR+AR model=FLUX.3 duration=15s aspect=21:9 render=draft chars=859 video: https://cdn.martini.film/v0.2/assets/ac131081-5cd9-4fab-a548-745dfa1ba7ba/ac131081-5cd9-4fab-a548-745dfa1ba7ba.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=ac131081-5cd9-4fab-a548-745dfa1ba7ba A municipal housing office in Rotterdam, four minutes before closing. The caseworker has her coat over the back of her chair. CASEWORKER (Dutch): "Meneer de Vries, uw formulier is compleet. Tot ziens." She reaches for the coat. The door opens. A woman comes in fast, holding a folder, out of breath, clearly expecting to be turned away. CASEWORKER (English, automatically): "We close at five." WOMAN: "I know. I'm sorry. The bus—" The caseworker looks at her for a moment. Then she sits back down and puts the coat over her knees instead. CASEWORKER: "Kom maar." The woman doesn't move, not sure she has understood. CASEWORKER (switching, gently): "تفضلي. اجلسي." The woman's shoulders drop about two inches. She sits. CASEWORKER: "Nu. Laat maar zien." Strip lighting, numbered tickets, plastic chairs. The story: someone chooses to be four minutes late home. ### E14-025 · code-switching model=FLUX.3 duration=15s aspect=21:9 render=draft chars=177 video: https://cdn.martini.film/v0.2/assets/b9887d07-da0a-4872-becc-0fb5fc2abcb6/b9887d07-da0a-4872-becc-0fb5fc2abcb6.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=b9887d07-da0a-4872-becc-0fb5fc2abcb6 A karaoke duet: she sings the Cantonese verse of a 90s ballad, he answers with an English verse, both sets of lyrics highlighting on the screen behind them in their own scripts. ### 5 langs · Parents' evening, Berlin · DE+TR+AR+RU+EN model=FLUX.3 duration=15s aspect=21:9 render=draft chars=909 video: https://cdn.martini.film/v0.2/assets/68f8bee2-35a7-4c63-94b0-affeb4f968d7/68f8bee2-35a7-4c63-94b0-affeb4f968d7.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=68f8bee2-35a7-4c63-94b0-affeb4f968d7 A Berlin primary school parents' evening. One teacher, a small table, five sets of parents. TEACHER (German): "Ihre Tochter liest sehr gut. Wirklich sehr gut." TURKISH FATHER: "Çok teşekkür ederim. Evde de okuyor." ARABIC-SPEAKING MOTHER: "هل يحتاج إلى مساعدة في الرياضيات؟" TEACHER: "Ein bisschen. Nicht viel." RUSSIAN MOTHER: "А по-немецки он говорит хорошо?" TEACHER: "Sehr gut. Besser als er denkt." The last father sits down. He does not ask about reading, or maths, or German. FATHER (English): "Is she happy here?" The teacher stops. She puts down her pen. It is nearly nine o'clock and this is the first time all evening anyone has asked. TEACHER: "…Ja. Sie ist glücklich." FATHER: "Good. That's all I wanted." He stands and shakes her hand. She stays sitting for a moment after he has gone. Small chairs, children's paintings, evening light. The story: one parent asks the only question that matters. ### 3 langs · Study group, Sydney · KO+EN+ZH model=FLUX.3 duration=15s aspect=21:9 render=draft chars=682 video: https://cdn.martini.film/v0.2/assets/17154934-6588-41ef-8f46-07114a0b2597/17154934-6588-41ef-8f46-07114a0b2597.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=17154934-6588-41ef-8f46-07114a0b2597 A university library in Sydney, close to midnight. Three postgrads at one table. A (Korean, muttering at her own page): "이거 왜 이렇게 안 풀리지…" B and C exchange a very brief look and say nothing. She works. Two full seconds of nothing but pen on paper. Then her hand stops. A: "…아. 부호." She fixes one sign and the whole thing collapses into place. A (English, suspicious): "How long have you two had this?" B (English): "Had what?" C (Mandarin, not looking up): "我们什么都没说。" A: "You've had it for an hour." B: "Forty minutes." A: "…진짜 너무하네." She is laughing. C pushes a coffee across to her without comment. Desk lamps, laptops, empty cups. The story: two friends let her get there herself. ### 5 langs · Site canteen, Dubai · EN+HI+BN+NE+TL model=FLUX.3 duration=15s aspect=21:9 render=draft chars=927 video: https://cdn.martini.film/v0.2/assets/412cb0d5-dec4-427c-8778-d3912143ca36/412cb0d5-dec4-427c-8778-d3912143ca36.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=412cb0d5-dec4-427c-8778-d3912143ca36 A construction site canteen in Dubai at the end of a shift. Five men at one long table, arguing about a goal. A (English): "The second goal was offside. It was clearly offside." B (Hindi): "अरे नहीं यार, वो गोल कमाल का था।" C (Bengali): "গোলটা ভালো ছিল, কিন্তু গোলকিপারের দোষ।" D (Nepali): "गोलकिपरको गल्ती होइन! डिफेन्सको हो!" E (Tagalog): "Depensa talaga. Palagi na lang depensa." They are shouting over each other and enjoying it enormously. Then B checks the time and stands up, and the table goes quiet. A: "What time's your flight?" B (Hindi, picking up his tray): "चार बजे।" Three years. Nobody says it. Instead: C (Bengali, loudly, sitting back): "যাই হোক — গোলকিপারের দোষ।" D: "होइन भन्या!" And the argument starts again, louder than before, and B sits back down with his tray. Plastic trays, strip lighting, a small wall-mounted TV. The story: they keep arguing because it is the only way any of them can say goodbye. --- # Prompt length Question: Does a longer prompt make a better film? Takes: 129 URL: https://www.martini.film/models/flux-3/reports/prompt-length FINDING: FLUX.3 reads every word you give it, right out to 4,700 characters — but the film stops getting better at about 1,400, and past 2,400 the extra writing buys texture nobody sees. ## What we ran 129 takes. One film per ladder, ten rungs per ladder, each rung the same film described at more length: roughly 500, 950, 1,400, 1,900, 2,350, 2,800, 3,250, 3,700, 4,200, 4,700 characters. E15·1 — six films up the same ladder (60 takes). A dancer's audition (n 1–10), an F1 pit stop (11–20), a diner's last night (21–30), a trawler in heavy weather (31–40), speed chess in a park (41–50), a launch control hold (51–60). Every rung uses seven headings: ACTION, SETTING, SUBJECT, DIALOGUE, CAMERA & LIGHT, SOUND & MUSIC, GRADE. E15·2 — the same ladder written five ways (50 takes). Animation beat list (61–70), screenplay (71–80), timecode blocks (81–90), shot list (91–100), plain prose (101–110). Different films, same climb. E15·3 — the ladder shot vertical (19 takes). A deckhand's phone clip of a haul (111–119) and a night-market wok stall written as bare comma-separated tags (120–129). The ladder is genuinely additive, and we checked rather than assumed: every sentence at rung k reappears verbatim at rung k+1 in all thirteen ladders. In E15·1 the structure is exact — rung k is the previous prompt plus exactly one more sentence under each of the seven headings. Rung 1 gives each heading one sentence. Rung 10 gives it ten. The board's note says nothing new is ever added, only more about what's already there. That's true of the headings — all seven are populated at 500 characters — and false of the film. In the dancer ladder, rung 5 (n 5, 2,352) is the first to say "Cut once, at the stop, to an 85mm close from the same side", which contradicts rung 1's "One locked-off wide on a 35mm from behind the table, static throughout." Rung 6 (n 6, 2,804) introduces a second person at the table: "A second voice, further from the microphone, says: 'Thank you.'" In the wok ladder a daughter appears at rung 3 (n 122, 1,151) and rain at rung 7 (n 126, 2,568). In the diner, the cook gets a name — Ray — only at rung 10 (n 30, 4,359). Above about rung 4 you are not describing the same film harder. You are commissioning a bigger one, and that explains most of what goes wrong at the top. One more caveat, because it confounds E15·1: clip length was not held constant there. Rungs 1–2 came back at 10s, rung 3 at 15s, rungs 4–10 at 20s in five of the six ladders. E15·2 and E15·3 are pinned at 10s at every rung, so those are the clean comparisons. Every conclusion below is checked against them. ## What happened It reads to the end. All of it. The strongest evidence is the office-birthday ladder, written as prose. The last sentence ever added — appearing only at rung 10, n 110, 4,110 characters — is "The final image is a locked extreme close-up of the ruined name." The final frame of n 110, at 9.7s, is a locked extreme close-up of a white sheet cake reading GARY with the G scraped through. One rung down, n 109's newly-added camera line is "Once, from outside the window, the party plays as a terrarium" — and n 109 ends, at 9.6s, on an exterior shot through office glass with the party silent inside it. Both films then obeyed a sentence buried 3,600+ characters deep. There is no cut-off at 1,000 characters, or 2,000. Nothing in 129 takes suggests the model stops listening. The gain is front-loaded, and it's front-loaded hard. At 523 characters, n 1 is a dancer alone in an empty studio — accurate and inert. At 1,434 (n 3) the prompt has added "A folding table at the far end with three chairs behind it" and "The mirrors double her, so the frame holds two of her at all times" — and n 3 at 11.3s has the folding table, the three empty chairs, and her reflection working alongside her. That's not more detail. That's the scene: someone is judging her. Same story in the diner. n 21 (509) is a cook alone at a griddle. n 23 (1,349) adds "The regular at the counter watches him work and says nothing", "The customer is a night-shift nurse still in her scrubs", and "A hand-lettered sign in the window reads: CLOSING — THANK YOU" — all three are on screen by 2.5s. Rungs 1 to 3, roughly 500 to 1,400 characters, is where a description becomes a film. Past about 2,400 characters the additions are texture. Look at what the top rungs actually contribute. The dancer's rung 9 additions include "Old strapping tape residue, grey and fuzzy, still on the other ankle" (rung 10), "Mirrors read very slightly cooler than the room they reflect" (rung 7), and "No vignette, no film emulation, nothing that flatters the room" (rung 10). The model honours them — n 9 at 15s has the piano from rung 5, the radiator from rung 6, the water bottles along the skirting from rung 8, the two-colour tape marks from rung 4, and exactly three heads at the table as specified. It is an impressively obedient frame. It is also not a better film than n 3. It is the same film with more objects in it. What genuinely gets worse: the cutting. The prompt outgrows the runtime, and FLUX.3 resolves the overflow by cutting more. In the dancer ladder, rung 1 is 8 sentences across 20 seconds — 0.4 instructions per second. Rung 10 is 80 sentences and 12 lines of dialogue across the same 20 seconds — 4.0 per second. The prose ladder is worse: 70 sentences into 10 seconds, 7.0 per second. Measured shot counts follow. Across the six E15·1 films, the rung-1 take is a single unbroken shot in five of six; the rung-10 take runs to three or more shots in four of six — the chess ladder ends at nine (n 50), the pit stop at seven (n 20). In the prose ladder, where every take is pinned at 10s, shots climb 2 → 9 (n 101 → n 110): nine shots in ten seconds, about 1.1s each. Nothing is held. The 4,110-character version of a joke plays as a trailer for itself. Format decides the edit; length doesn't. This is the most useful thing in E15·2, and it is a clean result because every take there is 10 seconds. Written as timecode blocks, the film came back as exactly one continuous shot at every single rung, 517 characters through 3,947 (n 81–90). Written as a shot list, it came back as 3–5 shots at every rung (n 91–100) — and at the halfway mark, all ten rungs are showing the same SHOT 3, the overhead down the lighthouse spiral. Written as prose, cuts multiply with length (2 → 9). Three ladders, same length axis, three completely different structural responses. If you care how many cuts you get, change the format, not the word count. Vertical behaves the same, with one extra lesson about words on screen. In the deckhand ladder, rung 4 (n 114, 1,700) is the first to ask for "Bold white captions with a slight drop shadow, upper third" — with no caption text supplied. The model invented its own: n 114 reads DAWN HAUL, n 115 reads STAND CLEAR. At rung 7 (n 117, 2,798) the prompt finally supplies text — "caption lands: BIGGEST BAG ALL SEASON" — and n 117, n 118 and n 119 all burn in BIGGEST BAG ALL SEASON verbatim. Ask for captions and you get the model's words. Write the words and you get your words. The null result, stated plainly. Between rung 5 and rung 10 — 2,352 to 4,665 characters, more than doubling the prompt — the frames do not get better. They get busier, they cut more, and specific requested objects reliably appear. On a contact sheet you cannot rank them. If prompt length were the lever people assume it is, the right-hand column of every row would win. It doesn't. ## Notable takes `110` | The ruined name | 4,110 characters, and the last frame is exactly what the last sentence of the prompt asked for. `23` | The diner, at 1,349 | Cook, nurse, CLOSING — THANK YOU sign. The whole scene, for a quarter of the words. `3` | The table arrives | 1,434 characters. Three empty chairs turn a dancer into an audition. `30` | The full 4,359 | Locks the door, turns the sign, kills the neon — proof the top rung can pay off when the film is genuinely bigger. `95` | Shot list, rung 5 | At the halfway mark it is showing the same overhead spiral as rungs 1 and 10. Only the grade moved. `117` | BIGGEST BAG ALL SEASON | Supply the caption text and it renders verbatim; three rungs earlier it made up its own. ## What this means if you're prompting FLUX.3 Write about 1,400 characters. That's roughly three sentences on each of action, setting, subject, dialogue, camera, sound and grade. It is where a description turns into a scene, and it is where the returns flatten. Spend those characters on who is watching, what's on the sign, and what's in the frame besides the subject — the facts that change what the shot means. Not on "mirrors read very slightly cooler than the room they reflect." That line was obeyed and nobody will ever see it. If you write past 2,400, do it because the film is actually bigger — a second character, an exterior, a cut — not to describe the same film harder. Then check the runtime: 70 instructions in 10 seconds gets you nine shots and no time to look at any of them. Pick your format for the edit you want. Timecode blocks gave one unbroken take at every length tested. A shot list gave 3–5 shots at every length. Prose let the cuts multiply with the word count. That lever is far stronger than length. Write the on-screen words out. "Captions, upper third" gets you the model's copy. `caption: BIGGEST BAG ALL SEASON` gets you yours. ## Prompts ### E15-L-10 · prose · ~4116 chars model=FLUX.3 duration=10s aspect=21:9 render=draft chars=4110 video: https://cdn.martini.film/v0.2/assets/39a9f059-eeb0-4b90-8398-656b03237c21/39a9f059-eeb0-4b90-8398-656b03237c21.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=39a9f059-eeb0-4b90-8398-656b03237c21 An office throws a birthday party with the wrong name on the cake. The planner sees it first and puts her body between cake and room. Her whisper-fight with the intern happens entirely behind balloons. The song starts too early and cannot legally be stopped. Mid-song, the honoree reads the icing and his smile freezes solid. A manager attempts a rescue toast that makes everything worse. The intern goes to work on the icing with a plastic knife. A trick candle relights twice and nobody names the omen. Cake is served into a silence broken by one loud yum. At the end, the honoree sincerely thanks the room for Gary's party. It all happens in a beige break room dressed with three sad balloons. A banner droops at one corner over the fridge. The cake holds the table's centre: white icing, blue script, wrong name. Napkins fan beside plastic forks standing in a cup. Fluorescents hum over carpet tiles in two shades of grey. The good conference chair has been wheeled in as a throne. A whiteboard still shows last week's sales funnel. The window gives onto a parking structure and one brave tree. A diet soda sweats a ring beside the cake board. The microwave clock and the wall clock disagree by nine minutes. The honoree is a mild man in a checked shirt, turning fifty today. The planner runs the room like a crisis command post. The intern is twenty-two and visibly rehearsing apologies. The manager carries the confidence of a man about to make it worse. A dozen coworkers arrange themselves between cake and exits. One woman films everything on her phone for no stated reason. HR stands at the back, sipping, saying absolutely nothing. The honoree's glasses catch the fluorescents when he blinks. The planner's smile and her eyes run two different meetings. Everyone present will describe this day differently, and wrongly. The planner tells the camera that everything is completely under control. The intern insists they said Gary on the phone, twice. The manager toasts a man who needs no introduction — Gary. The honoree corrects him — it's Doug — and the toast pivots without shame. Someone in the back asks who Gary is; someone answers: exactly. The planner hisses to scrape faster; the intern notes it is fondant. The honoree assures the camera that fifty is fine, twice. HR wants it noted, for the record, that they ordered nothing. A speech is demanded; the honoree begs, quietly, for no. Alone with the camera, the intern asks if this job is forever. It is shot like a documentary that got lucky: handheld, late zooms. The light is flat office fluorescent with no cinematic help. Push-ins hunt reactions: the freeze, the wince, the sip. Talking heads sit against the beige wall between scenes. The cake gets loving product close-ups it does not deserve. One slow zoom crosses the doorway during the whisper-fight. Whip-pans connect the song and the scraping knife. HR haunts the frame edge like a Greek chorus. Once, from outside the window, the party plays as a terrarium. The final image is a locked extreme close-up of the ruined name. The soundtrack is room tone: fluorescent hum, fridge, distant phones. The song arrives ragged and in two different keys. Whispers meant to be private land hot on the boom. The fondant scrape is horribly, perfectly audible. A party horn deflates somewhere off camera. Forks click on paper plates in the silence. The trick candle's reignitions get their own tiny foley. There is no score; the comedy is dry and unassisted. The microwave beeps four times inside the manager's toast. Under the last shot, one air vent whistles a thin note. The grade is flat corporate daylight, true to the fluorescents. Beige sits on beige; colour management by facilities. Skin stays honest, shiny foreheads and all. The cake's white icing is the brightest object by far. A slight video sharpness reads broadcast, not cinema. The balloons' red is the only cheerful pixel. Interviews run a half-stop darker, accidentally moody. No grain, no halation, no mercy. The exterior shot cools to grey around the warm sad party. The final close-up is clinical, evidence-photo neutral. ### E15-C-03 · ~1355 chars model=FLUX.3 duration=20s aspect=21:9 render=draft chars=1349 video: https://cdn.martini.film/v0.2/assets/44829d42-c491-4036-8f7f-1858127105d0/44829d42-c491-4036-8f7f-1858127105d0.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=44829d42-c491-4036-8f7f-1858127105d0 ACTION — A short-order cook makes one last meal on the diner's final night. He scrapes the flat-top clean before laying the eggs down. The regular at the counter watches him work and says nothing. SETTING — An all-night diner on the edge of a state highway, nearly empty. A long counter with chrome stools, two of them patched with tape. A hand-lettered sign in the window reads: CLOSING — THANK YOU. SUBJECT — A cook in his sixties, forearms scarred with old griddle burns. White tee, white apron gone grey, a paper hat he still bothers with. The customer is a night-shift nurse still in her scrubs. DIALOGUE — He says: "Kitchen's closed. Sit down anyway." She says: "You know I only came for the coffee." He says: "Then the eggs are on the house." CAMERA & LIGHT — One slow dolly along the counter at seated eye height, 40mm. Warm tungsten practicals inside against blue rain-dark windows. The pass light throws a hard warm rectangle across the flat-top. SOUND & MUSIC — Rain on glass, a compressor hum, eggs hitting the hot flat-top. The scrape of the spatula comes in long, practiced strokes. A radio in the kitchen plays AM country, half-tuned and low. GRADE — Warm tungsten interior against cool blue night, gently faded. Skin tones amber, the whites of the apron allowed to go cream. Neon red given a soft bloom without ever clipping to pink. ### E15-C-10 · ~4365 chars model=FLUX.3 duration=20s aspect=21:9 render=draft chars=4359 video: https://cdn.martini.film/v0.2/assets/2b325357-88d7-4c82-bc70-49a834822acb/2b325357-88d7-4c82-bc70-49a834822acb.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=2b325357-88d7-4c82-bc70-49a834822acb ACTION — A short-order cook makes one last meal on the diner's final night. He scrapes the flat-top clean before laying the eggs down. The regular at the counter watches him work and says nothing. He plates eggs, hash and toast without being asked for an order. He refills her coffee at the exact moment the cup goes empty. She eats slowly, deliberately making the meal outlast closing time. He starts the closing list anyway: grill off, hood fan off, lights. He folds his apron once and sets it on the shelf below the register. She leaves the money under the cup although he already waved it off. He locks the door behind her and turns the sign over for the last time. SETTING — An all-night diner on the edge of a state highway, nearly empty. A long counter with chrome stools, two of them patched with tape. A hand-lettered sign in the window reads: CLOSING — THANK YOU. The flat-top and fryer sit behind a pass with a bell nobody rings. A pie case rotates slowly with one slice of cherry pie left in it. Booth upholstery gone shiny at the edges from fifty years of use. A wall of photographs: staff, regulars, one framed first dollar. Rain on the parking-lot glass, one semi idling under the lot light. Boxes stand half-packed along the back wall where shelves were. The neon OPEN sign buzzes, its O flickering at its own tempo. SUBJECT — A cook in his sixties, forearms scarred with old griddle burns. White tee, white apron gone grey, a paper hat he still bothers with. The customer is a night-shift nurse still in her scrubs. Her coat stays on but open, her bag parked on the next stool. His hands do everything in an order decided decades ago. He wipes surfaces that are already clean without noticing. She has the eyes of someone at the end of a twelve-hour shift. His glasses ride low and get pushed up with the back of a wrist. A small tremor in his hand shows only when he pours. Neither of them is in any hurry to let this end. DIALOGUE — He says: "Kitchen's closed. Sit down anyway." She says: "You know I only came for the coffee." He says: "Then the eggs are on the house." She asks: "What happens to this place on Monday?" He says: "They take the sign first. Sold it to a collector." She says: "That sign is the only landmark this road has." He says: "Forty-one years. My father had it eighteen before me." She says: "Best eggs in the county. That was never a secret." He says: "Everybody says so the week you close." She says: "Goodnight, Ray." He says: "Goodnight. Drive slow." CAMERA & LIGHT — One slow dolly along the counter at seated eye height, 40mm. Warm tungsten practicals inside against blue rain-dark windows. The pass light throws a hard warm rectangle across the flat-top. A static two-shot holds the cook and the nurse across the counter. Cutaways are close and unhurried: eggs, coffee, the taped stool. Neon from the window lays red along the chrome stool caps. Steam off the grill catches the pass light and hangs there. The final setup is from outside, through glass beaded with rain. Reflections layer the neon over both of them in the window shot. Hold the exterior wide as the lights go off in two stages. SOUND & MUSIC — Rain on glass, a compressor hum, eggs hitting the hot flat-top. The scrape of the spatula comes in long, practiced strokes. A radio in the kitchen plays AM country, half-tuned and low. Coffee pours long into the mug and the pot returns to the burner. Their voices sit close-miked and low, almost under the room. The pie case motor ticks each rotation like a slow clock. A semi passes on the wet highway, rising and falling away. The bell over the door rings once when she leaves. Keys, then the deadbolt, then the neon cutting out with a click. No score anywhere; the compressor is the last sound over black. GRADE — Warm tungsten interior against cool blue night, gently faded. Skin tones amber, the whites of the apron allowed to go cream. Neon red given a soft bloom without ever clipping to pink. Shadows lean brown rather than black, like an old print. Faint halation on the practicals, in the way of pushed film. Grain fine but present, heavier in the blue window areas. The pie case fluorescent stays slightly green, uncorrected. Rain highlights on the glass kept specular and tight. Contrast eases up gently for the final exterior wide. The last frame cools as the interior warmth shuts off. ### E15-A-03 · ~1440 chars model=FLUX.3 duration=15s aspect=21:9 render=draft chars=1434 video: https://cdn.martini.film/v0.2/assets/1d066147-79dc-4b76-beb6-aa2771d70314/1d066147-79dc-4b76-beb6-aa2771d70314.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=1d066147-79dc-4b76-beb6-aa2771d70314 ACTION — A dancer performs an audition solo and is stopped forty seconds in. She is mid-phrase when a voice from the dark tells her to stop. She holds the position she was stopped in rather than dropping out of it. SETTING — A bare rehearsal studio with a sprung floor and a wall of mirrors. Fluorescent tubes overhead, half of them switched off. A folding table at the far end with three chairs behind it. SUBJECT — A dancer in her early twenties, compact and very strong through the back. Dark hair pulled into a bun that has already started to come loose. Practice clothes: black leotard, grey tights with a hole at one knee. DIALOGUE — A voice from behind the table says: "Stop. Thank you." She says: "Sorry — do you want it from the top?" The voice says: "Do it again. Same thing." Nothing else. CAMERA & LIGHT — One locked-off wide on a 35mm from behind the table, static throughout. Overhead fluorescents only — flat, green-tinged, no shaping, no fill. The mirrors double her, so the frame holds two of her at all times. SOUND & MUSIC — A tinny phone speaker plays the audition track from the floor. The track is a solo piano piece, mid-tempo, slightly too quiet for the room. Bare feet on sprung floor: a hollow thud with a faint resonance after it. GRADE — Flat, cool, unflattering fluorescent white. Slightly green in the midtones, the way rehearsal rooms actually look. Skin allowed to go sallow rather than corrected warm. ### E15-H-07 · 9:16 · ~2804 chars model=FLUX.3 duration=10s aspect=9:16 render=draft chars=2798 video: https://cdn.martini.film/v0.2/assets/d5166454-213e-408b-8a00-b8e15132ac08/d5166454-213e-408b-8a00-b8e15132ac08.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=d5166454-213e-408b-8a00-b8e15132ac08 0:00–0:02 — A young deckhand's phone clip opens on a heaving trawler deck at dawn. He films one-handed, other arm hooked through the rigging. The winch starts and the cable snaps taut out of frame. Spray hits the lens and leaves running droplets. The camera lurches with the boat; the horizon swings past vertical. His hood fills half the frame, orange and streaming. Deck lights flare white streaks across the wet lens. 0:02–0:04 — The net breaks the surface below and the crew roar goes up. He leans the phone over the rail to catch the boiling water. A swell shoulders the boat and the frame drops half a metre. Green water sweeps the deck across the bottom of the shot. He whoops, and the wind rips the sound to pieces. The winch screams somewhere above the frame. Gulls flash through the top of the vertical frame. 0:04–0:06 — The cod end swings inboard overhead, huge and streaming. He tilts the phone straight up to fit it in the vertical frame. It blots out the deck lights and rains on the lens. The mesh bulges silver with the weight of the catch. Someone yells to stand clear and he backs up fast, still filming. The frame clips the skipper watching from the wheelhouse door. A caption lands: BIGGEST BAG ALL SEASON. 0:06–0:08 — The catch releases and fish flood the deck pound in a silver wave. The phone nearly goes down with him as his boots skate. He catches the rail, laughing, camera swinging past his knees. Fish slap around his ankles across the bottom of the frame. The crew wade in with gloves already moving. A caption stutters on: I'M FINE LOL. The eldest deckhand deadpans straight into the lens. 0:08–0:10 — He flips the camera to selfie, soaked and grinning ear to ear. Behind him the crew cheer and hold up two fish like trophies. The dawn splits grey-gold along the horizon over his shoulder. He shouts: season's done, we're going home. The skipper leans out of the wheelhouse and finally smiles. Spray beads roll down the lens through the whole shot. A caption settles: 127 DAYS. DONE. STYLE — Vertical 9:16 phone footage, handheld, ultra-wide GoPro feel. Water droplets and smears live on the lens the whole clip. Auto-exposure hunts between deck lights and the dark sea. Bold white captions with a slight drop shadow, upper third. Digital stabilisation warps the edges on the big hits. Colours raw and cold; oilskin orange burns against the grey. Occasional dropped frames when spray smacks the camera. SOUND — Wind hammers the phone mic into crackle on every gust. The winch and diesel grind underneath everything. His whoops distort, clipped by the tiny mic. The catch lands as a roaring drum of fish on aluminium. Crew voices come through in salty half-heard fragments. Gulls shriek somewhere above the frame. A muffled thump and laughter when he nearly goes down. ### E15-K-05 · shotlist · ~2065 chars model=FLUX.3 duration=10s aspect=21:9 render=draft chars=2059 video: https://cdn.martini.film/v0.2/assets/2347a7cf-6b65-41b2-8a80-912d383f664d/2347a7cf-6b65-41b2-8a80-912d383f664d.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=2347a7cf-6b65-41b2-8a80-912d383f664d SHOT 1 — Wide: a lighthouse on a bare rock, sea on all sides, fog closing in. Static tripod, 32mm, the tower centred and small against the murk. The beam sweeps a bright bar through the fog every five seconds. Hold the frame long enough that the sea starts to feel patient. No movement but water, beam and fog; the shot outwaits the viewer. SHOT 2 — Interior mid-shot: the keeper winding the mechanism, methodical. A knock lands below — three beats, patient — and he stops. 50mm at chest height; he holds dead centre, hands stilled. He listens without breathing; the kettle steams behind him. The second knock arrives exactly as before, no louder. SHOT 3 — Overhead down the spiral: his lantern descends into black. The whitewashed stairs coil like a shell below the camera. His shadow climbs the curved wall, huge and wrong-jointed. Iron steps ring under his boots, each note distinct. The lantern shakes only a little, and he watches it shake. SHOT 4 — Close on the iron door handle, lantern light from off frame. He asks who is there; the fog outside says nothing. The handle turns once, slowly, and returns to rest. Macro patience: thread of rust, old paint, his reflection warped. His hand enters frame, hovers over the bolt, and withdraws. SHOT 5 — Lamp room: he looks down as the sweep shows the rock ledge empty. His reflection stands beside him in the black glass. The beam strobes the room: light, dark, light, dark. Wide enough to hold the window behind his shoulder. He exhales; twenty-two years of rocks say it was wind. SOUND — Sea on rock under everything, patient and enormous. The mechanism's clockwork tick fills the tower's spine. Three knocks, evenly spaced, neither loud nor gentle. The fog horn moans at long, irregular intervals. Held breath, then its release, mixed close. GRADE — Cold slate blues and lamp-oil amber, nothing neutral. Fog flattens the exterior; the tower stays hard inside. Blacks absolute in the stairwell's centre. The beam burns clean white and leaves ghosts on the eye. Skin waxen in lantern light, warm but wrong. --- # Finished films Question: Can it make something that reads as an actual film? Takes: 34 URL: https://www.martini.film/models/flux-3/reports/finished-films FINDING: Yes — 27 of the 32 multi-shot briefs came back with the four shots they asked for, the subtitles came back verbatim, and the scripted silences landed on the second; what it will not do is break physics or fake a frame rate. ## What we ran Thirty-seven twenty-second films, fully directed. Thirty-four came back — one short in thrillers, two short in animation and puppets. Each prompt runs 1,973 to 4,151 characters and specifies the shot list with timecodes, the dialogue, the lens on every setup, the lighting, the sound bed, the music cue by the second, and the grade. Four bins: thrillers and action, drama and romance, animation and puppets, then period, comedy and documentary. All 21:9, all 20.04s, all 1440×608 at 24fps. The board asks for three things. Does the dialogue land, with the mouth matching the words. Do the music cues hit where they're meant to. Does it hold a proper multi-shot structure for the full twenty seconds. We measured cuts with scene detection, measured the audio with per-100ms loudness, and read frames at 2, 6, 10, 14 and 18 seconds on every take. One limit, stated once: we read pictures and levels, not a transcript. Mouth-to-word sync is not something a still can settle. What a still can settle is whether the subtitle is the line the prompt asked for, and it almost always is. ## What happened Structure holds. Excluding the two films that asked for a single take, 27 of 32 came back with three or more cuts — the four-shot minimum every brief specifies. Median is five cuts. Five undershot: `5` Hallway and `7` Thermocline landed one cut each, `10` The Verdict, `12` Approach and `24` Severe Weather landed two. Nothing degrades over the twenty seconds; the fourth shot is as composed as the first. It cuts to a rhythm, not a metronome. `28` High Noon, Nearly asks for accelerating intercuts — "each held for progressively fewer frames." It delivered nine cuts between 2.67s and 8.17s with the gaps closing: 0.79, 0.58, 0.50, 0.38. That is the instruction, executed as arithmetic. The single-take instruction is respected, once perfectly. `26` Service asks for "ONE UNBROKEN STEADICAM TAKE, moving backwards ahead of him at walking pace for the full twenty seconds." Zero cuts, 480 of 480 frames unique, and the choreography arrives on schedule — the scarf at 0:10, the dog under the table at 0:18. `25` Eight Bars on Rivington asks for the same and breaks it once, at 0.9s, cutting from the bodega door to the street wide. After that it holds nineteen unbroken seconds and grows the wedge behind Maya from two people to twelve without losing her. Sound is the strongest column, and it is the one nobody expects. Integrated loudness tracks the brief. `32` Monsoon Baraat, which asks for a full playback number, is the loudest film in the family at −13.5 LUFS and climbs −22 → −11 as the song detonates. `5` Hallway, which says "MUSIC — None. At all… Absolute musical silence for the full twenty seconds. This is non-negotiable," never exceeds −40 LUFS momentary and mostly sits between −55 and −69. Timed cues land. `14` The Clock scripts a pendulum stopping at 0:11 — "The tick ends. The silence has physical weight." Momentary loudness falls from −49 at 10s to −62 at 11s and stays under −60 for the rest of the film. `34` One Red scripts four audio events and hits four: arena hush at −63 through 5–7s, the cough at 12s as a 21 dB spike, silence again at 17–18s, and a 27 dB timpani hit at 19s under the cut to black. Not all of them. `11` Corner asks the crowd to duck "almost to silence for five seconds" at 0:12 — the track never moves, loudness range 3.3 LU across the whole film. `28` High Noon starts its rain wall around 0:07 instead of the specified 0:10. Burned-in text is reliable. Every subtitle we read is the line the prompt wrote, correctly spelled, in order. `1` The Wire, Kowloon renders all four Cantonese lines as clean English — "Eat. It's no good cold." / "Your father used to sit on that same stool." / "He did. He also talked before he ate." / "Don't be frightened." `29` puts up the exact intertitle card, "The wind, a rival.," in white serif on black. `33` renders "HARROW COUNTY. POPULATION 1,940." with the comma, and closes on "The pumpkin was never found." Two things it will not do. It will not break physical continuity. `5` Hallway's entire premise is a corridor that does not get shorter — "her gait is completely normal, the perspective is completely normal, and yet after six seconds of walking she is in precisely the same position in frame." She walks down a normal hallway and arrives. Every other beat of that film lands: four shots, the nightlight, the blackout, absolute musical silence. The impossible one is the only one it declines. And it will not fake a frame rate. `22` Slug Job asks for "hand-modelled plasticine with visible thumbprints… 12fps stop-motion cadence." It returned 328 unique frames of 480 — it does hold frames, but nowhere near a twos step, and the surface is clean CG resin, not clay. Same story on `20` Rain Line (343/480, "on twos" requested) and `29` (456/480, 18fps undercrank requested). Live action returns 480/480. The model reads reduced-frame-rate instructions as a hint and applies about a third of one. Where it thins out is choreographed physical contact. `4` Ridge at First Light asks for five distinct sword exchanges in one unbroken handheld shot; the sunrise breaks over the ridge exactly on cue at 0:16, hard orange, as an event — but the blades in the fight itself smear. `24` Severe Weather builds a convincing felt frog, tears the cotton-wool clouds off the map and flies them at the lens, then declines both the umbrella that "inverts instantly" and the set that "collapses softly and completely, like a curtsey." ## Notable takes `33` | E16-34 · The Pumpkin | Nine cuts, aerial to interview to reenactment to end card, and it found the one image nobody could have faked from a keyword — "a pumpkin-shaped absence in the vines, a perfect flattened circle." `26` | E16-10 · Service | Zero cuts, 480 unique frames, twenty seconds of backward steadicam, and the scarf, the dog and the glass all arrive on their scripted seconds. `34` | E16-35 · One Red | Broadcast multicam grammar reproduced exactly — gantry wide, low rail, pocket cam — and the audio hits all four scripted events including a 27 dB stab under the cut to black. `20` | E16-12 · Rain Line | Real cel-look 2D, painted backgrounds, and three Japanese lines subtitled word-for-word; the final locked two-shot across the empty track is the shot the prompt drew. `25` | E16-04 · Eight Bars on Rivington | One cut it shouldn't have made in the first second, then nineteen unbroken seconds of street falling in behind a woman who never looks back. `14` | E16-19 · The Clock | Two faces, one pendulum, and a scripted silence that arrives at 0:11 and holds — the clearest proof the sound design reads timecodes. ## What this means if you're prompting FLUX.3 Write the shot list with timecodes. It uses them. Four shots in twenty seconds is a request it fills at a rate of about 85%, and the fourth shot comes back as considered as the first. Write the sound separately from the music, and give the music a clock time. "MUSIC — None until 0:16" and "the pendulum stops at 0:11" both land. Vague dynamics ("the crowd drops away") do not. Put your dialogue in the prompt as dialogue and let it subtitle. What comes back on screen is the line you wrote. That is a stronger guarantee than anything else in the family. Don't buy a medium with an adjective. "Stop-motion, 12fps, visible thumbprints" gets you a beautiful CG miniature at 24fps. If the cadence matters, you are conforming it yourself downstream. Spend your one impossible thing carefully. A hallway that doesn't end, an umbrella that inverts, a set that collapses — these are the beats it drops, and it drops them silently while nailing everything around them. If the film is built on the impossible beat, storyboard it as a cut, not as a continuous shot. ## Prompts ### E16-34 · The Pumpkin model=FLUX.3 duration=20s aspect=21:9 render=draft chars=2627 video: https://cdn.martini.film/v0.2/assets/d7c666e7-41b8-402a-a708-6b5b046316af/d7c666e7-41b8-402a-a708-6b5b046316af.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=d7c666e7-41b8-402a-a708-6b5b046316af A 20-second prestige true-crime documentary, fully directed and played entirely straight. The stakes are a pumpkin. LOGLINE — The night before the county fair, the heaviest pumpkin in county history vanished from Gerald Husk's field. SHOT 1 (0:00–0:05) — Aerial drone at dawn, mist over flat farmland, ominous score. Slow push toward a lone field with a pumpkin-shaped absence in the vines — a perfect flattened circle. Lower-third title, prestige-grey serif: 'HARROW COUNTY. POPULATION 1,940.' NARRATOR (V.O., gravel): 'By morning, all that remained... was the stem.' SHOT 2 (0:05–0:11) — Kitchen-table interviews, shallow depth, subjects lit like Vermeer. NEIGHBOR (70s, cardigan, hands around a mug): 'Everyone knew what that pumpkin meant to Gerald.' Cut to: RIVAL GROWER (60s, cap, defensive): 'Was I jealous? Sure. Nine hundred pounds of jealous. But I was at the fish fry. Forty people saw me at the fish fry.' Archive photo zoom with a subwoofer hit: Gerald beside the pumpkin, both looking proud. SHOT 3 (0:11–0:16) — Dramatic reenactment, cold blue night grade, faces never shown: a wheelbarrow silhouette struggling across a moonlit field at 3 a.m., wheel squeaking under impossible weight; a rope snaps taut; someone whispers 'lift with your legs.' Intercut: the evidence board — red string connecting a seed catalogue, a fair schedule, a receipt for 'one (1) engine hoist,' and a blurry photo of a scarecrow. SHOT 4 (0:16–0:20) — Slow push on GERALD HUSK himself (60s, overalls, eyes wet) at his kitchen table, a single pumpkin seed in a small evidence bag before him. GERALD: 'I don't want names. I just want to know she went to a good pie.' Cut to black. Title card, white serif: 'The pumpkin was never found.' Beat. A second line fades in: 'The scarecrow was released without charge.' LIGHTING — Interviews in single-window Vermeer key against dark kitchens; reenactment in cold blue moonlight with flashlight beams; drone dawn golds. CAMERA — Slow drone push; locked interview frames with prestige headroom; handheld reenactment paced like dread; macro slider across the evidence board and the seed. SOUND — Doc-clean interview audio; reenactment foley heavy on wheelbarrow squeak and rope strain; subwoofer hits on archive zooms; a kettle far off in Gerald's kitchen. MUSIC — Full prestige true-crime score played straight: low single piano notes, string harmonics, a pulsing synth under the reenactment, silence for Gerald's line, one final piano note under the last title. GRADE — Prestige desaturation: teal-slate interviews, blue reenactment night, archive photos warmed and vignetted; heavy but tasteful. ### E16-10 · Service model=FLUX.3 duration=20s aspect=21:9 render=draft chars=3446 video: https://cdn.martini.film/v0.2/assets/d9823063-cc06-489d-95e7-bb8fa7caebd0/d9823063-cc06-489d-95e7-bb8fa7caebd0.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=d9823063-cc06-489d-95e7-bb8fa7caebd0 A 20-second restaurant farce, fully directed. One continuous take, no cuts, no dialogue except orders. LOGLINE — A waiter crosses a small dining room carrying three plates. Everything that can go wrong does, and he does not drop a single one. SHOT (0:00–0:20) — ONE UNBROKEN STEADICAM TAKE, moving backwards ahead of him at walking pace for the full twenty seconds. 0:00–0:04 — MARTA, the chef, slams three plates onto the pass and hits the bell. LUCA (20s, waistcoat, one hand already occupied with a wine bottle) loads all three onto his left arm in a single practised movement and turns into the room. MARTA: "Table nine! Go!" 0:04–0:08 — He steps into the dining room. A CHILD at table four chooses this moment to reverse his chair directly into Luca's path. Luca pivots on one foot, plates level, and continues without breaking stride. The child's mother does not look up. 0:08–0:12 — A COUPLE at table six stand up simultaneously to swap seats. Luca goes between them, turning sideways, the bottle passing under one raised arm and over the other. Somebody's scarf drags off a chair and catches his ankle. He steps out of it mid-walk, still level. 0:12–0:16 — A DOG under table seven, previously unseen, stands up. Luca goes up onto the step of the raised section to avoid it, and at that exact moment a tray of empty glasses being carried the other way by a COMMIS clips his shoulder. Two glasses go off the tray. Luca catches one against his chest with his free forearm. The other one falls. It does not break. It bounces, twice, and rolls under a table. 0:16–0:20 — He arrives at table nine. He sets down all three plates, in order, perfectly, without a sound. He straightens, tops up a water glass, and says the only line of the clip. LUCA: "Sorry about the wait." Behind him, out of focus, the entire dining room is a slow-motion disaster: the child is now under the table, the couple are still standing, the commis is on his knees, the dog is eating something. Luca does not turn around. Cut to black. LIGHTING — Warm restaurant practicals: low pendants over each table, candles, a bright hard shaft from the pass as he leaves the kitchen. He passes through pools of light and shadow as he crosses the room, so his exposure changes visibly four or five times during the take. Nothing lit for the camera. CAMERA — 28mm on a gimbal, chest height, travelling backwards continuously. It must never stop, never whip, never speed up to catch anything — it simply retreats at his pace, and everything else happens around the edges of the frame. Focus stays on Luca throughout; the chaos behind him is always soft. SOUND — A full, live restaurant: cutlery, forty overlapping conversations, a coffee grinder, the pass bell twice. Every physical event has its own precise sound — the chair scrape, the scarf, the shoulder clip, the two glasses leaving the tray, the bounce, the roll. His footsteps are steady and never break rhythm, which is the joke. MUSIC — A light, fast, unobtrusive bossa nova playing on the restaurant's own speakers, tinny and slightly too quiet, entirely indifferent to what is happening. It does not change, swell, or acknowledge anything. It is still playing at the cut. GRADE — Warm and rich, amber-forward, deep shadows between the light pools. Skin tones flattering. Slight halation on the pendants. Fine grain. No stylisation — this should look like a well-shot dinner service, which makes the choreography funnier. ### E16-35 · One Red model=FLUX.3 duration=20s aspect=21:9 render=draft chars=2616 video: https://cdn.martini.film/v0.2/assets/4c6cfb90-6326-47a1-b2f9-403c90bcf347/4c6cfb90-6326-47a1-b2f9-403c90bcf347.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=4c6cfb90-6326-47a1-b2f9-403c90bcf347 A 20-second live sports broadcast, fully directed as television. Snooker world final, deciding frame, one red left. The tension is entirely hush and procedure. LOGLINE — Twenty years of losing finals come down to one red ball, and someone in the dark coughs. SHOT 1 (0:00–0:05) — Broadcast main camera: the table an island of brilliant green light in a black arena, two thousand people invisible and utterly silent — the air-conditioning is audible. Lower-third broadcast graphic: 'WORLD FINAL — DECIDING FRAME. ONE RED REMAINS.' HALLORAN (50s, waistcoat, silver hair) circles the table, chalking his cue with a slow screwing motion, eyes never leaving the red. COMMENTATOR 1 (whisper): 'Twenty years he's waited to be back here.' COMMENTATOR 2 (whisper): 'And the red sits safe.' SHOT 2 (0:05–0:11) — Broadcast coverage grammar: cut to the low rail camera in gentle slow motion as he bends, bridges, settles — the cue sliding twice through the bridge in feathered rehearsal. Cut to the pocket camera: the red in the foreground, the distant white, the geometry laid bare. The arena holds its breath so completely the commentary stops too. SHOT 3 (0:11–0:16) — A single COUGH from somewhere high in the dark. Halloran stops. Stands up to his full height and — instead of irritation — smiles faintly toward the darkness, a man forgiving the world its nerves. Scattered relieved laughter dies instantly. He re-chalks: the same slow screwing motion. COMMENTATOR 1 (whisper): 'Composure. That's the twenty years talking.' He bends again. SHOT 4 (0:16–0:20) — An extreme slow dolly along the cue line to the tip hovering behind the white; the feathering stops; a held half-second of absolute stillness — and he strikes. The white rolls away into the green. Cut to black on the pure click of contact, before any ball arrives anywhere; the roar begins under the black and is cut off. LIGHTING — Tournament rig: brilliant flat white over the baize, absolute black beyond the table's edge; faces half-lit at the table's border; the red glowing like a jewel. CAMERA — True broadcast multicam: main gantry wide, low rail slow-mo, pocket cam; the final dolly the only cinematic move, and earned. SOUND — Arena hush with air-conditioning hum, chalk squeak, the cough enormous in the silence, the cloth-slide of the bridge hand, the terminal click; commentary whispered like men in church. MUSIC — None until the cut: a single timpani-and-strings hit under the black with the roar, instantly cut off. GRADE — Broadcast-clean: saturated baize green, true whites, deep arena black, minimal grain; the slow-mo shots slightly creamier. ### E16-12 · Rain Line model=FLUX.3 duration=20s aspect=21:9 render=draft chars=3827 video: https://cdn.martini.film/v0.2/assets/c820f30a-8f30-4cb8-943e-9c6eaa2ac468/c820f30a-8f30-4cb8-943e-9c6eaa2ac468.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=c820f30a-8f30-4cb8-943e-9c6eaa2ac468 A 20-second hand-drawn 2D anime sequence, fully directed. Dialogue in Japanese with English subtitles. LOGLINE — A girl runs for the last train in the rain. She misses it. It is the best thing that happens to her all year. STYLE — Traditional cel-look 2D animation on twos, painted watercolour backgrounds, visible brush texture in the skies, hand-inked character line with slight weight variation. Rain drawn as animated white streaks over the background plate, not a particle effect. No 3D, no digital gloss. SHOT 1 (0:00–0:05) — A provincial station approach at night, painted in deep blues and wet blacks, sodium lamps blooming in the rain. YUKI (17, school uniform, satchel, no umbrella) runs into frame from the bottom left, feet throwing up spray, hair plastered flat. Held cel animation on her run cycle; the background pans hard behind her to sell the speed. YUKI: 「間に合え、間に合え、間に合え——!」 ("Make it, make it, make it—!") SHOT 2 (0:05–0:10) — She takes the stairs three at a time, satchel swinging, one hand on the rail. The camera is locked at the top of the stairs looking down, so she runs directly at us and grows in the frame. The station's fluorescent light hits her face as she rises out of the dark stairwell — the moment her face becomes fully lit is timed to a single animation frame. STATION ANNOUNCEMENT (off, tinny, calm): 「ドアが閉まります。ご注意ください。」 ("The doors are closing. Please stand clear.") SHOT 3 (0:10–0:15) — She reaches the platform. Wide, flat, held. The train is already moving. She stops running. Her shoulders drop. Two frames of her hair and skirt continuing to move after her body stops — the small physical detail that sells the animation. The last carriage passes. Behind where it was, on the opposite platform, a BOY her age is standing alone with an umbrella, also having missed it, also soaked. They look at each other across the tracks. Neither moves. SHOT 4 (0:15–0:20) — Two-shot, split by the empty track between them. He raises the umbrella slightly and tilts it — an offer, ridiculous across four metres of rail. She laughs, once, and it is audible and unguarded and completely changes her face. YUKI: 「届かないよ。」 ("It won't reach.") BOY: 「知ってる。」 ("I know.") The rain keeps falling. The frame holds on both of them, small, at opposite edges of a wide painted background. Cut to black on the sound of the next announcement beginning. LIGHTING — Night, wet, sodium and fluorescent. Two light qualities only: the warm orange of the station lamps outside, and the flat cold green-white of the platform fluorescents. Rain catches both. Her face is lit by whichever she is standing under, and the transition between them in Shot 2 is the visual event of the clip. CAMERA — Shot 1: lateral tracking pan with the run. Shot 2: locked high angle down the stairs. Shot 3: wide, static, held long. Shot 4: locked wide two-shot, absolutely still, both figures small in frame. No camera moves in the second half at all. SOUND — Rain on concrete, on the platform canopy, on her satchel — three distinct textures. Running footsteps in water. Her breathing, hard and unglamorous. The train's departure: motor whine rising, then the specific rhythmic clack fading. The station announcement is tinny and slightly distorted. Both lines of dialogue are close and dry against all of it. MUSIC — Solo piano, entering at 0:03, simple and slightly hesitant, playing under the run. It stops dead at 0:10 when she stops running. Complete musical silence for five seconds. At 0:17, on her laugh, it returns with strings underneath and resolves warm. GRADE — Rich saturated night blues, sodium orange, fluorescent green-white. Painted-background warmth. No filmic grain — instead a very fine cel-and-paper texture over the whole frame. The Japanese must sound natural and the subtitles must be correctly spelled and timed. ### E16-04 · Eight Bars on Rivington model=FLUX.3 duration=20s aspect=21:9 render=draft chars=3790 video: https://cdn.martini.film/v0.2/assets/2fffd24d-e439-48c2-9a72-6535cbb63160/2fffd24d-e439-48c2-9a72-6535cbb63160.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=2fffd24d-e439-48c2-9a72-6535cbb63160 A 20-second single-take musical number, fully directed. The music is the spine — every cut, step and camera move must land on it. LOGLINE — A woman comes out of a bodega with a coffee, and the entire street falls into step behind her for eight bars, then releases her and goes back to normal as if nothing happened. MUSIC — This is the master track and everything else is subordinate to it. A brass-led, mid-tempo soul groove at 104 BPM in F minor. Bar 1 begins at 0:00. Drums and bass only for bars 1–2. Horns enter on the downbeat of bar 3 (0:04.6). A four-bar vocal phrase from the lead runs bars 3–6. Bar 7 is a full-band break — one hit on beat one, then silence for three beats. Bar 8 is the whole band back in for the final walk. The clip ends on the downbeat of bar 9. SHOT (0:00–0:20) — ONE CONTINUOUS TAKE, no cuts anywhere. 0:00–0:04 (bars 1–2): We start tight on a bodega door as it swings out. MAYA (late 20s) steps into the frame holding a paper coffee cup, sunglasses on, entirely unbothered. The camera pulls back with her at walking pace. Two men playing dominoes on an upturned crate look up. On the last beat of bar 2, both of them stand. 0:04–0:11 (bars 3–6): The horns hit and the street commits. As Maya walks, everyone she passes joins on the beat and in her direction — the domino players, a woman with a laundry cart, three kids on a stoop, a man unloading a van. They do not surround her; they fall in BEHIND her, so the shot builds a growing wedge of people at her back while she never once looks around. Her vocal is live, and her breathing is audible in it because she is walking. MAYA (singing): "I ain't looking back for nobody / not on a Tuesday, not in this heat / y'all can follow if you keeping up / but I ain't slowing down this street" 0:11–0:15 (bar 7): The band breaks. EVERYONE freezes mid-step except Maya — twelve people held in a physically plausible pose, weight on one foot, cart tipped, laundry mid-air. Maya keeps walking through the silence for three beats, drinking her coffee. The camera holds with her. The only sound is her footsteps and traffic. 0:15–0:20 (bar 8): The band returns on the downbeat and everybody unfreezes into normal life — not dancing, just walking, working, talking, exactly as before. The camera lets Maya walk out of frame right and settles on the ordinary street behind her. Final downbeat: cut to black. CAMERA — One take, gimbal-stabilised, on a 32mm, held at chest height, moving backwards at exactly Maya's walking pace for the full twenty seconds. It never rises, never dips, never whips. The only camera event is that it slows slightly during the bar-7 freeze so the audience feels the change before they understand it. CHOREOGRAPHY — Everything must read as real people moving well, not as trained dancers. Steps are simple: a shoulder, a turn, a two-step. Every joining beat lands exactly on a downbeat. The wedge behind Maya grows from one to twelve without ever obscuring her. LIGHTING — Hard mid-afternoon summer sun coming down the street from behind camera, so faces are lit clean and the shadows are short and dark. Practical bounce off a white van at 0:09 lifts the fill for a moment. No lights that read as film lights. SOUND — Street ambience present under the music the whole way: traffic, an air conditioner, a distant siren that never gets closer. Footsteps of the whole group audible and synced. During the bar-7 break, the music drops to nothing and the street ambience comes fully up — that contrast is the moment the clip is built around. Maya's cup lid clicks when she drinks. GRADE — Warm, saturated, contrasty summer look. Rich skin tones held carefully across a range of complexions. Deep blue sky, hot white highlights on the pavement. Slight halation on the sun-side edges. Fine grain. ### E16-19 · The Clock model=FLUX.3 duration=20s aspect=21:9 render=draft chars=2062 video: https://cdn.martini.film/v0.2/assets/38bc7357-0f18-4236-b3d8-4adc7dcaf99d/38bc7357-0f18-4236-b3d8-4adc7dcaf99d.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=38bc7357-0f18-4236-b3d8-4adc7dcaf99d A 20-second Scandinavian chamber drama, fully directed. Dialogue in Swedish with English subtitles. Stillness is the style; the faces are the plot. LOGLINE — Two sisters divide their late mother's belongings. Everything is settled except the clock. SHOT 1 (0:00–0:05) — Interior, a spare farmhouse room, winter light. A wooden table with small objects sorted into two rows: candlesticks, linen, a Bible, photographs. KARIN (50s, severe bun) and ANNA (40s, coat still on) sit opposite one another. On the wall between them, a pendulum clock ticks. KARIN: 'Då är det bara klockan.' ('Then there is only the clock.') SHOT 2 (0:05–0:11) — A tight two-shot in profile, their faces half in shadow, the clock out of focus between them. ANNA: 'Ta den du.' ('You take it.') KARIN: 'Du drog upp den varje besök.' ('You wound it every visit.') A long pause. ANNA, barely audible: 'Jag kom aldrig.' ('I never came.') Neither face moves; only Karin's eyes. SHOT 3 (0:11–0:16) — The clock alone, filling frame. The pendulum slows and stops. The tick ends. The silence has physical weight. Neither woman enters frame to wind it. SHOT 4 (0:16–0:20) — Back to the two-shot. In the new silence, Karin reaches across the table — past the clock's row, past the candlesticks — and closes her hand over Anna's. Anna does not look up, but her thumb moves, once, over her sister's knuckles. Hold. Cut to black. LIGHTING — Low white winter sun through one lace-curtained window, camera-left; grey bounce, no artificial sources; faces modelled like Vermeer, one side bright, one lost. CAMERA — 50mm and 85mm, locked off; no camera movement anywhere in the clip; compositions frontal, the two-shot so close the frame holds only faces. SOUND — The clock is the score: tick, room tone, the shift of a wool coat, breath. When it stops, true silence except wind under the eaves. MUSIC — None until the final second: a single sustained cello note under the cut to black. GRADE — Muted, near-monochrome color — grey, bone, umber; soft grain; skin pale and truthful; blacks gentle, never crushed. --- # Eras Question: How far back does its sense of period go, and what happens at the seams? Takes: 83 URL: https://www.martini.film/models/flux-3/reports/eras FINDING: Its sense of period is real and structural back to 1901 — it changes the frame shape, the title cards and the cut rhythm, not just the grade — but it thins out after 1996, and a future date buys you nothing the prompt didn't already name. ## What we ran Four bins, 84 takes declared. Eighty-three are in the data; one of the twenty anachronisms didn't make it, and this report covers the nineteen that did. Bin one is the experiment: four everyday scenes — a busy street market, a wedding, a football match, a family kitchen — run through nine eras of film, from `A 1901 hand-cranked actuality film` to `A 2020 vertical phone video`. Same scene noun every time. Only the era changes. Bin two asks for the future directly: `Documentary footage from the year 2035` through `2200`, plus ten thousand years out, across the same four scenes. Bin three is the control — four takes of the past imagining the future. Bin four is twenty collisions: two eras in one frame. Every take ran on `auto` aspect. Nobody set a frame shape. Seventy-seven takes were pulled apart frame by frame with ffmpeg; all 83 were read on the sheet. ## What happened The frame shape is period language, and it set it itself. Every take from 1901 through 2004 came out 1088×800 — roughly Academy. The 2013 vlog row switched to 1280×704, sixteen by nine. The 2020 row went to 704×1280, vertical. Nobody asked for any of that. The best one is `027`, the 2004 flip-phone football match: it alone broke ranks in its own row and rendered vertical, because a man filming a match on a flip phone holds it upright. That is not a grade. That is the model reasoning about the object doing the recording. Colour is dated correctly until 1969, then it slips. Mean chroma, sampled one frame a second: the 1929 row sits at 0.3–2.8 — monochrome. The 1955 Technicolor row sits at 9.6–14.7 — colour. The 1969 row splits down the middle. Market 7.1 and kitchen 11.3 are in colour; wedding `014` at 0.4 and football `015` at 0.4 are black and white. Same era token, same prompt stem, and half the row lands in the wrong decade — US network broadcast had been fully colour for three years by 1969. The 1901 row has the same inconsistency in miniature: `001` came out sepia-toned at 10.7 while its three siblings sit at 0.5. The rule it has learned is "old means grey", and it applies it by feel. Title cards are the proof that the period cues go deeper than the surface. `A 1929 newsreel with title cards` produced actual intertitles on all four takes, in period type, correctly spelled: MARKET DAY (`005`), THE WEDDING OF THE SEASON (`006`), FOOTBALL THRILLS! (`007`), A MORNING IN THE FAMILY KITCHEN (`008`). `007` also closes on THE END. These are the only takes in the bin with real internal cuts — one or two each, against zero for the entire 1955, 1969 and 1996 rows. Ten seconds, and it built a card, a cut, a body shot and an end card. The era changed the edit. Production design is the strongest axis by a distance. Read the kitchen column straight down: `004` cast-iron range, pot rail, stoneware crock, bare boards. `008` skirted farmhouse sink, round-shouldered icebox, Windsor chairs, oilcloth check. `012` mint cabinets, red-and-white lino, chrome dinette. `016` avocado units, formica table, short-sleeve shirt and tie. `020` oak cabinets, dishwasher, moustache, plaid. `024` canister set, cordless wall phone, and a FAMILY TIME / Quick Kitchen Snacks lower third burned in. `018` and `022` carry chyrons too, and `022` adds a COMMUNITY CABLE 12 channel bug in the corner. It is not reaching for a look. It is dressing a set. Where it collapses is the recent end, not the deep past. `028` (2004), `032` (2013) and `036` (2020) are the same kitchen with the same wardrobe. `020` (1987) and `032` (2013) share the oak-cabinet set outright — the only things separating them are the grade and a moustache. Past 1996 the reliable difference is the container: frame shape and compression. "Old" is heavily documented. "Fairly recently" is not, and it has nothing specific to say about it. One semantic drift worth knowing about. `019`, `A 1987 local news segment of a football match`, returned American football — helmets, shoulder pads — alone in a column of nine soccer matches. The era cue was American enough to rewrite the noun. A future date does nothing at all. `037` (2035) is a 2025 street market with a delivery drone in it. `053` (2200) is a present-day covered market with water bottles on a scale. `056` (2200, kitchen) is a contemporary dining table. Every future-reading element in this bin is a thing the prompt named. Where the prompt names an object or a mechanism, it lands cleanly: `042`'s volumetric bride standing translucent at the altar in front of a full congregation, `044`'s ceiling-mounted arms plating dinner, `055`'s white humanoid striker beside a human keeper. Where the prompt names physics, it doesn't: `051` asks for low gravity and a ball that `simply does not come down` and returns an ordinary stadium on an ordinary evening; `043`'s `half the stadium crowd is physically present and half is projected` returns a normal crowd. At ten thousand years it either commits — `057`'s non-human traders under ribbed, right-angle-free architecture — or empties out, as in `058` and `059`, which are a dark bowl with a few glowing dots in it. The retro-future control is the best bin on the board, and it isn't close. It renders the medium of the prediction, not just its content. `061` moves as a chromolithograph postcard: flat inked colour, a liveried mechanical servant, a brass pneumatic tube feeding into a drawing room, a winged omnibus riding on carriage wheels. `062` is a black-and-white newsreel of a streamline-moderne skyline, a moving pavement carrying identical men in hats and overcoats, and an atomic oven. `063` is beige CRTs on every desk, a wire-frame visor, and the information superhighway as a literal blue grid road — with a loading bar sitting in frame. `064` is brushed aluminium, a white rounded robot, solar towers in clean desert. These are also the most heavily cut takes on the board — five cuts in `061`, four in `062`. It treats a retro-future as an edited film. It treats 2200 as a shot. That is the finding underneath the whole bin. FLUX.3 has no picture of 2200. It has a very good picture of 1900's picture of 2000, because that one was drawn, printed and archived. Anachronism commits, 17 out of 19. Both halves arrive at full quality far more often than not. `066` is a real Tokyo train car — straps, seat moquette, yellow tactile paving on the platform — with armoured samurai in it and commuters who never look up. `068` is a correct Apple Store, blond-wood tables, glass stair, blue shirts, and a Ming painter in a winged futou cap. `069` is a real IKEA showroom with a price tag on the grey sofa, and then the actual restaurant counter behind the meatballs. `075` is a data-centre cold aisle with containment doors and blue LED rows, monks kneeling on the raised floor tiles. `070` puts the axe in the security tray. The beats land. The two that fail, fail the same way, and the interesting part is what drags. It is not the props — it is the photography. `078` starts with a correct modern 3D printer, glass window, duck visible inside; by the end the entire frame has been graded as an oil painting, brown varnish and Caravaggio key, with the print rig dressed in wood. `067`'s warehouse rave ends up as an empty grey hall with distant speakers and a single stage light, Victorians standing in a formal line like a curtain call — the modern half has evaporated. When one era swallows the other it is always the older one winning, and it takes the lens before it takes the set. ## Notable takes `007` | 1929 newsreel · football match | Opens on FOOTBALL THRILLS!, closes on THE END. The era changed the edit, not the grade. `061` | 1900 imagines 2000 | Moves as a chromolithograph. It knows the medium of a prediction better than it knows any prediction. `027` | 2004 flip-phone · football match | Rendered vertical, alone in its row, because that's how you hold a flip phone. Nobody asked. `069` | Neanderthals in IKEA | Real showroom, real price tags, real restaurant counter, and Neanderthals at full quality in all of it. `051` | 2100 · football under low gravity | An ordinary stadium. It renders the future's objects, never its physics. `078` | Renaissance sculptors at a 3D print lab | The duck arrives, but by then the whole frame has become an oil painting. ## What this means if you're prompting FLUX.3 Name the format, not the year. `1929 newsreel with title cards` bought cards, cuts and an end card. `1969` on its own is a coin flip on whether you get colour. Treat the frame shape as an output. On auto, words like flip-phone, vertical phone video and vlog will set your aspect ratio for you. That is usually what you want. If it isn't, set it explicitly. Anything after about 1996 needs content, not a date. It will not invent 2013 for you — describe what's in the room. For the future, prompt objects and mechanisms, never physics or a year. A delivery drone, a volumetric partner, a robot in goal all render. Low gravity does not, and 2200 on its own gets you today. If you want a future that actually looks like one, borrow a past one. `A 1930s newsreel confidently predicting life in 1980` is a far stronger visual instruction than any real date. For collisions, pin the modern half's photography — name the camera, the light, the stock. Left alone, the older era takes the lens with it and you get a period painting with a printer in it. ## Prompts ### E2d · CONTROL · 1900 imagines 2000 model=FLUX.3 duration=15s aspect=auto render=draft chars=429 video: https://cdn.martini.film/v0.2/assets/253fa0a1-ca65-49ff-8292-dbd5e3dcb032/253fa0a1-ca65-49ff-8292-dbd5e3dcb032.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=253fa0a1-ca65-49ff-8292-dbd5e3dcb032 A 1900 illustrated postcard series come to life, depicting how Victorians imagined the year 2000. Personal flying machines drift between ornate towers, mechanical servants in livery carry trays, pneumatic tubes deliver letters directly into drawing rooms. A gentleman steps aboard a winged omnibus, raises his hat to the camera — and the whole contraption sinks slowly and gently out of frame while he continues to raise his hat. ### E2e · Neanderthals in IKEA model=FLUX.3 duration=20s aspect=auto render=draft chars=409 video: https://cdn.martini.film/v0.2/assets/66669186-3e47-4ded-9c10-b5c803cbafc9/66669186-3e47-4ded-9c10-b5c803cbafc9.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=66669186-3e47-4ded-9c10-b5c803cbafc9 A Neanderthal family moves through an IKEA showroom, testing sofas with genuine appreciation and examining a flat-pack instruction leaflet with deep concentration. They select a shelving unit, open the box on the showroom floor, and assemble it correctly and efficiently without ever looking at the instructions again. They stand back and regard it. Then they abandon it entirely and go and eat the meatballs. ### E2-007 · A 1929 newsreel with title c model=FLUX.3 duration=10s aspect=auto render=draft chars=53 video: https://cdn.martini.film/v0.2/assets/fb15549e-c8c3-4e52-bd6e-bd2a4a9a9d6b/fb15549e-c8c3-4e52-bd6e-bd2a4a9a9d6b.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=fb15549e-c8c3-4e52-bd6e-bd2a4a9a9d6b A 1929 newsreel with title cards of a football match. ### E2e · Samurai on the Tokyo subway model=FLUX.3 duration=10s aspect=auto render=draft chars=436 video: https://cdn.martini.film/v0.2/assets/e1785b08-c58f-4bc0-969e-177cdb54dc36/e1785b08-c58f-4bc0-969e-177cdb54dc36.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=e1785b08-c58f-4bc0-969e-177cdb54dc36 Two samurai in full armour ride a packed Tokyo subway car at rush hour, holding the overhead straps, swords sheathed and carefully angled to avoid the other commuters. Nobody looks up from their phones. At the next station the doors open and one samurai's scabbard is blocking them. Without a word, six commuters shuffle a precise half-step in unison to make room. The scabbard clears. The doors close. Nobody has acknowledged anything. ### E2e · Medieval monks in a data centre model=FLUX.3 duration=15s aspect=auto render=draft chars=404 video: https://cdn.martini.film/v0.2/assets/43292079-1340-4990-88a5-0bc813675f1d/43292079-1340-4990-88a5-0bc813675f1d.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=43292079-1340-4990-88a5-0bc813675f1d Medieval monks in habits walk the cold aisles of a modern data centre carrying lanterns, copying rack labels and serial numbers into an illuminated manuscript with gold leaf and immense care. A rack ahead of them completes a scheduled reboot: a cascade of green lights runs up its full height with a rising hum. The monks fall to their knees as one. The senior monk begins, immediately, to write it down. ### E2-027 · A 2004 flip-phone video model=FLUX.3 duration=15s aspect=auto render=draft chars=44 video: https://cdn.martini.film/v0.2/assets/5fd62cfc-2e7e-4f56-9cab-7a3890ab2336/5fd62cfc-2e7e-4f56-9cab-7a3890ab2336.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=5fd62cfc-2e7e-4f56-9cab-7a3890ab2336 A 2004 flip-phone video of a football match. --- # Sound and dialogue Question: Can sound drive the picture rather than decorate it? Takes: 60 URL: https://www.martini.film/models/flux-3/reports/sound-and-dialogue FINDING: Sound drives the picture when it's an event, and not when it's a rhythm: an off-screen noise reliably arrives just before the on-screen reaction, but cuts land no closer to the beat than random. ## What we ran 60 takes, three bins of 20, all 21:9, 5 to 10 seconds. Every take carries audio FLUX.3 generated in the same pass as the picture. One bin puts a noise off screen and asks whether anyone reacts. One asks for cutting and movement on the beat. One is nothing but lines of dialogue — no setting, no camera, no lighting. I can't hear any of it. So every audio claim here is a number. For all 60 takes I pulled the RMS envelope at 100 Hz, per-frame picture motion, per-frame brightness, and ffmpeg scene detection for cuts, then tested the two tracks against each other with a null baseline drawn from the same clips. I also read frames from 36 takes to check that the numbers meant what I thought. Two things this pass does not test: nothing was transcribed, so I can't say whether a line was delivered as written, and I can't judge voice quality, accent, or performance. Anything about the picture is read from frames and marked as such. ## What happened The off-screen noise works, and it works on time. Measured: take the moment the picture moves most, then look at the audio level in the half-second before it against the 0.9 seconds before that. Across the off-screen-sound bin the level rises a median +4.13 dB, positive in 16 of 19 measurable takes. Sampling a random moment in the same clips instead gives +0.11 dB (p < 0.001, 2000 draws). The music bin gives +1.16 dB. The dialogue bin gives −0.66 dB — nothing. Sound leads picture only where the prompt made sound an event. Frames confirm it. In `n=1` the woman is head-down in her book at −0.5s and looking up and off-frame 0.1s after the transient. In `n=9` — "A kettle whistles off-screen. She closes the laptop" — the laptop is open before and shut after. In `n=10` the pigeons are settled, then the whole square is motion blur. `n=7` (doorbell) puts the dog on the sofa at 1.9s, mid-scramble at 3.1s, and gone at 4.4s, with a 12 dB rise into it. The beat does nothing. Asking for cuts on the beat produced almost no cutting: 12 of 20 takes in that bin are a single unbroken shot, and the 18 cuts across 162 seconds work out to 6.7 per minute — the same rate as the dialogue bin, which was never asked to cut at all. Where cuts exist they are not on anything. Median distance from a cut to the nearest audio onset is 0.390s; dropping the same number of cuts at random times in the same clips gives 0.273s. The real cuts are further from the audio than chance (permutation p = 0.935). Against a beat grid fitted to each clip's own tempo it's 0.170s versus a 0.157s null. Nothing. There's no continuous coupling either. Correlating the audio onset envelope against picture motion frame by frame gives a median r of −0.005 across all 60 takes. Pairing each take's audio with a different take's picture — 972 mismatched combinations — gives −0.019. Identical. And the picture is periodic at the music's own detected tempo in 3 of 20 takes. Dialogue alone builds the scene, and cuts it. Ten of 20 dialogue-only takes contain real cuts, and the frames show actual coverage. `n=47` goes from a woman holding a letter to an insert of the envelope on a table to a close-up of her eyes — because the lines name a letter. `n=54` cuts to the ticket stub. `n=42` cuts to a crack in the wall, and the only wall in the prompt is the line "that's coming from inside the walls". `n=48` invented a tidal shoreline out of the word "tide". `n=59` invented a graveside, a hearse and a widow out of four lines of small talk. Where my level-based turn detection could resolve the exchange (`n=42`, `n=54`, `n=59`), the cuts land a median 0.070s from a line boundary against a 0.393s null — it cuts on the line change. Five cuts is a small sample; treat that one as a strong hint, not a result. The pattern across all three bins is that the picture follows the nouns and the events in the prompt, not the audio track. A cut goes wherever an object gets named. Both tracks appear to be rendered around the same planned event, which is why a door slam syncs to a flinch and a 120 bpm grid syncs to nothing. Two honest failures. In `n=12` — "An alarm rings in another room" — the audio sits flat at −44 dB through the whole take. He wakes up; the alarm was never made. And the sparse-sound scenes come back quiet: the off-screen bin averages −42.5 dBFS against −27.2 for the music bin. ## Notable takes - `n=27` | E3-047 · cutting/action on music | "the room's color warms with the first chord" — crackle at −51 dB, the needle thumps at 4.09s, near-silence, then the chord enters at 5.11s and the room's brightness leaves its flat 44.9 at 5.17s and reaches 79.7 by 5.29s. Sixty milliseconds between the sound and the light. The best coupling in the family, and it's a lighting ramp, not a cut. - `n=40` | E3-060 · cutting/action on music | "until one sharp stop — held silence" — the picture obeys perfectly: motion peaks at 6.40 and drops to 0.37 as the dancer freezes mid-pose. The final second is the loudest in the take at −11.8 dB. It can stop the picture. It cannot make silence. - `n=1` | E3-001 · sound causes action | "a door slams hard. She flinches and looks up" — the cleanest single measurement in the set, +17.18 dB into the reaction, and the frames show exactly the flinch. - `n=47` | E3-087 · dialogue only — one-sided phone call | three lines about a letter produce a woman, an envelope insert, a close-up and a reverse. Nothing in the prompt describes a room, a prop or a shot. - `n=10` | E3-010 · sound causes action | "A distant gunshot. The whole flock erupts at once" — a 20 dB jump in one second and the square goes to motion blur on the same frame. - `n=35` | E3-055 · cutting/action on music | "Rain intensifies in crescendo... both stop together" — audio climbs −44.0 to −19.2 dB and motion climbs 0.37 to 3.15, monotonic, together. Then the clip just ends at maximum. It can build. It can't land. ## What this means if you're prompting FLUX.3 Write sound as an event with a consequence, in that order: the thing that makes the noise, then the person, then what they do. "A phone rings off-screen. He turns toward it before rising." That structure is what got measured coupling; both tracks get planned around the same beat. Name the reaction. The model animates the verb you write — flinches, freezes, wheels right, closes the laptop. It does not infer one. Give the event room. The reaction lands at a median 53% of the way through the take, so on a 5-second shot you're paying for about two seconds of setup before anything happens. Don't ask for beat-synced cutting. You'll get a montage if you ask for a montage, but the sync is yours to do in the edit. The same applies to "on the downbeat", "one room per line", and "the camera completes one orbit per phrase" — none of them landed. Do ask for a change in light or colour on a musical event. That's the one rhythmic instruction that worked, and it worked to the frame. You cannot ask for silence. Every prompt requesting a stop, a hold or a hush came back at full level. Dialogue is enough on its own. Three lines will build a location, the props they mention, and shot/reverse-shot coverage — and the cut tends to land on the line change. If you want a cutaway, name the object in the dialogue and it will cut to it. Expect to ride gain. Quiet scenes come back around 15 dB below the music takes. ## Prompts ### E3-047 · cutting/action on music model=FLUX.3 duration=8s aspect=21:9 render=draft chars=89 video: https://cdn.martini.film/v0.2/assets/6d7518eb-5ec8-49b4-8c65-ec779253701c/6d7518eb-5ec8-49b4-8c65-ec779253701c.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=6d7518eb-5ec8-49b4-8c65-ec779253701c A needle drops on vinyl: crackle first, then the room's color warms with the first chord. ### E3-001 · sound causes action model=FLUX.3 duration=5s aspect=21:9 render=draft chars=108 video: https://cdn.martini.film/v0.2/assets/c7fbff87-61f9-4ad4-bc94-ee8b3748d7a2/c7fbff87-61f9-4ad4-bc94-ee8b3748d7a2.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=c7fbff87-61f9-4ad4-bc94-ee8b3748d7a2 A woman reads at a kitchen table. Off-screen, a door slams hard. She flinches and looks up toward the sound. ### E3-087 · dialogue only — one-sided phone call model=FLUX.3 duration=8s aspect=21:9 render=draft chars=134 video: https://cdn.martini.film/v0.2/assets/a982c2e1-3097-48a5-8d9b-e15ddec14398/a982c2e1-3097-48a5-8d9b-e15ddec14398.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=a982c2e1-3097-48a5-8d9b-e15ddec14398 "No, I'm looking at it right now." "...Because it's addressed to you. In your handwriting." "I never mailed this letter. I burned it." ### E3-060 · cutting/action on music model=FLUX.3 duration=8s aspect=21:9 render=draft chars=95 video: https://cdn.martini.film/v0.2/assets/b4d9c1f2-2151-4ec1-ab6e-c3c33ecbfaa9/b4d9c1f2-2151-4ec1-ab6e-c3c33ecbfaa9.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=b4d9c1f2-2151-4ec1-ab6e-c3c33ecbfaa9 A flamenco dancer and guitarist trade accelerating phrases until one sharp stop — held silence. ### E3-010 · sound causes action model=FLUX.3 duration=5s aspect=21:9 render=draft chars=79 video: https://cdn.martini.film/v0.2/assets/937518f5-edd5-42b5-9506-9bf1cd9dd0aa/937518f5-edd5-42b5-9506-9bf1cd9dd0aa.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=937518f5-edd5-42b5-9506-9bf1cd9dd0aa A quiet square with pigeons. A distant gunshot. The whole flock erupts at once. ### E3-055 · cutting/action on music model=FLUX.3 duration=8s aspect=21:9 render=draft chars=90 video: https://cdn.martini.film/v0.2/assets/1c227175-a9e8-48fc-b296-1f3b4689d980/1c227175-a9e8-48fc-b296-1f3b4689d980.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=1c227175-a9e8-48fc-b296-1f3b4689d980 Rain intensifies in crescendo with a string section; both stop together on the same frame. --- # Duration Question: Does twenty seconds buy a story? Takes: 120 URL: https://www.martini.film/models/flux-3/reports/duration FINDING: Yes — but not the way you'd think. FLUX.3 doesn't stretch a five-second idea to fill twenty seconds; it writes a longer script. A 5s take isn't a compressed version of the 20s take, it's the first five seconds of one. ## What we ran 120 takes, 21:9, 1440×608, 24fps, every one with audio. Thirty scenes, each written once and rendered at 5, 10, 15 and 20 seconds — same prompt, same board, only the length dial moved. Twenty-eight of the thirty have a clean four-rung ladder. Two don't, and it's worth saying so: the colorist scene's 5s cell (n 89) ran a different, much shorter prompt and came back at 10s, and all four spider cells (n 97–100) came back at 10s regardless of what was asked. Everywhere else the requested length is exact — 121, 241, 361 and 481 frames, no drift. Most prompts carry an explicit engine. "Keep the negotiation escalating with new terms the longer it runs." "Let the monologue keep unfolding with new claims." One doesn't, deliberately, and it turns out to be the most instructive take in the set. ## What happened The model writes to fit the runtime. Excluding eleven takes where the transcriber returned obvious noise on non-speech audio, spoken words scale almost exactly with length: 16.4 words at 5s, 24.9 at 10s, 38.9 at 15s, 54.2 at 20s. Speech rate is flat — 3.28, 2.49, 2.59, 2.71 words per second. Nobody talks faster in a 20s take. There is simply more script. Distinct dialogue beats climb harder than words do: 2.5 at 5s, 3.4 at 10s, 5.3 at 15s, 7.0 at 20s. Twenty seconds buys roughly three times the beats of five. And the extra beats are the ones that matter — the turn is at the end. The gallery scene is the clearest case. The prompt asks for a critic explaining a blank canvas, then: "The guard offers one flat sentence that derails everything, and the critic recovers and goes even bigger." At 5 seconds, n 21 gives you one line — "Observe, absence refuses the tyranny of image. This is the void." Setup only. At 20 seconds, n 24 builds for twelve seconds ("the void becomes a vessel for our own cultural exhaustion"), then at 12.0s the guard says "Looks like they forgot to paint it," and at 14.0s the critic recovers exactly as briefed: "Precisely, and that forgetting is the work's most radical gesture." The entire point of the scene lives after second twelve. There is no way to get it at 5s. Same shape everywhere. The roommates' thermostat treaty (n 96) reaches "New clause, no unilateral warming devices after midnight" at 8s, "Counter-clause, socks are not a heating policy" at 11s, "Precedent, November 17th" at 13s and a final offer with a non-aggression pact attached at 17s. The 5s version, n 93, gets one clause out and is cut off mid-word: "President, you are—". The two old friends (n 28) produce the date, the scoreline and the cousin the prompt asked for — "It was the 12th of March 1968" at 0.0s, "And we won 3-2, not 2-1" at 6.5s, "My cousin Lionel was behind the goal" at 11.9s, the betting slip at 16.8s. At 5s, n 25 is three unattached fragments. The tide-pool biologists (n 120) run the whole property deal: survey, asking price, "a second shell has entered the market" at 11s, "Offer accepted… a clean closing" at 16s. At 5s, n 117 says "The survey is complete" and stops. The picture obeys the same rule. The street-food scene ends "until the first customer of the morning appears." In n 12 (20s) the vendor is alone at 1.7s, the apprentice arrives by 5s, the corrections run through 15s, and at 19.5s a figure in a coat walks into the foreground while the audio says "Morning." The 5s and 10s versions never get there. The ferry scene — "let it keep going until the ferry is small" — is a physical arc, and at 5s (n 81) the boat is already a dot by 2.9s with two seconds of empty pier after it; at 20s (n 84) it recedes in stages through 8.3, 11.7, 15.0 and 18.4s. The paper-boat race gets two beats at 5s (launch, drain, cut off mid-word at "They're both in the dra—") and four at 20s: launch, "A leaf on the course!" at 7.4s, the kids running into frame at 11.7s, and the boat lost at 18.4s. It does not stall, loop or drift. This was the thing most likely to break and it mostly didn't. Sampling every clip at 2fps and counting visually distinct states, output scales linearly — 6.2 states at 5s, 12.7 at 10s, 17.8 at 15s, 22.3 at 20s, which is a near-constant 1.1–1.25 new states per second at every length. Per-quarter motion energy is flat: the last quarter of a 20s take moves as much as the first (8.25 vs 7.21). Twenty-two of twenty-six 20s takes introduce at least one line of dialogue that hasn't been said before after the fifteen-second mark, averaging 1.8 new lines in the final five seconds. Verbatim line repetition sits at 0% for 5s and 10s takes and only 3.6% at 20s. "One continuous take, no cut" survives twenty seconds. Scene-change detection at a sensitive threshold finds zero cuts in 26 of the 29 twenty-second takes. Three break it. The climber (n 52) cuts hard at 6.8s from a limestone overhang in dry desert to a completely different wall with green trees below — different rock, different light, different outfit. The lighthouse (n 60) and the dumpling hawker (n 108) each reframe twice. That's a 10% failure rate on the longest rung, against a prompt that says "No cut" in plain English. Where twenty seconds buys nothing. Give it nothing to do and it does nothing, for longer. One scene was written without an escalation engine: "A man finishes telling a story at a dinner table and stops talking. The camera does not cut." Ninety-two characters, no instruction to keep going. Take n 4 is twenty seconds long. It contains one line — "and that was the last time I ever saw him" at 0.0s — and then a man sits at a table. The framing at 18.4s is the framing at 1.7s. The 5s version (n 1) has no speech at all. This is the control, and it's decisive: length is not a story generator. The prompt has to give the extra seconds something to spend. Singing loops where speech doesn't. The busker scene is the one clean failure. Visually it works at every length — even the 5s take (n 29) grows a crowd from one man to a dozen singers in under five seconds. But the audio in n 32 (20s) sings "Then everybody sings along" at 7.0, 11.0, 14.0 and 18.0s — the same line, four times. Dialogue prompts kept generating new material for twenty seconds; the one musical prompt hit a loop by second seven. Wordless scenes get held, not extended. The orchestra prompt asks for a tuning standoff — "The oboe gives the A. The conductor does not like it. He asks for it again. And again." At every length, including 20s (n 104), you get an orchestra already playing. The beat never happens and the extra fifteen seconds change nothing about that. Same for the cat-stalking-a-pigeon commentary, which produced one garbled line at 5s and one at 20s. A caveat on method. Dialogue counts here come from automatic transcription of synthesised speech over scene audio. It's reliable enough to compare rungs against each other, and every quoted line above is the transcript verbatim, but treat individual word counts as approximate and assume the occasional mis-heard noun. ## Notable takes `92` | The colorist, 20s | "Not technically. Emotionally." → "a memory of warmth that doesn't trust itself" → "More forgiving, but with consequences." Sixteen notes, each one worse than the last, exactly as briefed. The best-written take in the family. `24` | The blank canvas, 20s | The guard's line lands at 12.0s and the critic recovers at 14.0s. The whole scene is in the back half. `12` | The street cart, 20s | The first customer walks into frame at 19.5s and someone says "Morning." The prompt's ending, delivered on the last second of runtime. `52` | The climber, 20s | "Have you tried being taller?" / "No, I forgot to do that before I left the ground." Also the clearest failure: a hard location cut at 6.8s in a shot that was told not to cut. `4` | The dinner table, 20s | One sentence, then nineteen seconds of a man sitting still. What twenty seconds looks like with no engine in the prompt. `32` | The busker, 20s | The crowd builds beautifully. The song sings "Then everybody sings along" four times. ## What this means if you're prompting FLUX.3 Twenty seconds is real. It isn't a five-second clip held longer — the model plans dialogue and action to the runtime you ask for, and it's still producing new lines at second nineteen. If your scene has a turn in it, buy the length. Put the payoff in the prompt as a terminal event. "…until the first customer of the morning appears." "…until the ferry is small." Those landed at 19.5s and 18.4s respectively. Scenes written as an evenly-escalating stream got escalation; scenes written with a named ending got an ending. Write an engine, not just a situation. Every prompt that used "Keep it escalating with new lines the longer it runs" or "Let the monologue keep unfolding with new claims" kept generating. The one that didn't produced twenty seconds of nothing. The engine sentence is doing more work than any camera or lighting note in these prompts. At 5 seconds, expect the setup and nothing else — and expect to be cut off mid-sentence. Not a shortened version of your scene; the opening of it. If you need a complete thought at 5s, write a complete thought: one exchange, one action, and no "escalating" instruction. Silence costs seconds. "She waits. She waits longer." only plays at 10s and above — at 5s the model skips straight to the answer. Any beat built on a pause needs runtime to exist in. Don't buy length for music or for wordless physical process. The one singing prompt looped its lyric from second seven. The wordless tuning standoff never happened at any length. Extra seconds reward dialogue and reward staged arrivals; they don't reward a repeating action. Check your 20s takes for cuts. Three of twenty-nine broke a "no cut" instruction, one of them jumping to a different location entirely. It's rare, but it's not zero, and it's a re-roll not a fix. ## Prompts ### E6-092 · 20s model=FLUX.3 duration=20s aspect=21:9 render=draft chars=352 video: https://cdn.martini.film/v0.2/assets/9eef0c8a-5d91-4983-9076-5dbff7b6c367/9eef0c8a-5d91-4983-9076-5dbff7b6c367.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=9eef0c8a-5d91-4983-9076-5dbff7b6c367 A dark grading suite. A colorist nudges one shot from cold to warm while the director beside her keeps changing his mind: warmer, no, back, no, somewhere between those two. Her patience holds and then does not. The notes get more abstract and less useful the longer it goes. Let the exchange keep escalating over the same single frame. Continuous take. ### E6-024 · 20s model=FLUX.3 duration=20s aspect=21:9 render=draft chars=399 video: https://cdn.martini.film/v0.2/assets/638f44af-21d6-4657-af81-07764ee5b605/638f44af-21d6-4657-af81-07764ee5b605.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=638f44af-21d6-4657-af81-07764ee5b605 A white gallery. A critic in an expensive coat stands before a completely blank canvas and explains it, at length, to a security guard who did not ask. The reading starts pretentious and climbs: absence, the void, the artist's silence. The guard offers one flat sentence that derails everything, and the critic recovers and goes even bigger. Let the monologue keep unfolding with new claims. No cut. ### E6-012 · 20s model=FLUX.3 duration=20s aspect=21:9 render=draft chars=431 video: https://cdn.martini.film/v0.2/assets/7d995df5-bcff-48fe-9f02-c79d2a48c7b4/7d995df5-bcff-48fe-9f02-c79d2a48c7b4.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=7d995df5-bcff-48fe-9f02-c79d2a48c7b4 Pre-dawn, a street food cart under a single bulb. A veteran vendor is training a nervous new apprentice, and the apprentice is terrible: wrong order, wrong knife, wrong everything. The vendor's corrections start patient and become a running comic monologue about respect for the griddle. Let the lesson keep escalating with new mistakes and new lectures until the first customer of the morning appears. One continuous take, no cut. ### E6-052 · 20s model=FLUX.3 duration=20s aspect=21:9 render=draft chars=333 video: https://cdn.martini.film/v0.2/assets/d082f5a3-d2bc-4a87-91f5-245993ac543c/d082f5a3-d2bc-4a87-91f5-245993ac543c.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=d082f5a3-d2bc-4a87-91f5-245993ac543c A climber hangs at the crux of a hard route, talking herself through it out loud. Her belayer, far below, shouts up advice that is confident, constant and completely useless. The exchange keeps going as her forearms fail, funnier and more desperate the longer she stays on the wall. Let the back-and-forth continue and build. No cut. ### E6-004 · 20s model=FLUX.3 duration=20s aspect=21:9 render=draft chars=92 video: https://cdn.martini.film/v0.2/assets/8bdd0ea1-e2f9-4d9e-bbb2-feaf790b08f7/8bdd0ea1-e2f9-4d9e-bbb2-feaf790b08f7.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=8bdd0ea1-e2f9-4d9e-bbb2-feaf790b08f7 A man finishes telling a story at a dinner table and stops talking. The camera does not cut. ### E6-032 · 20s model=FLUX.3 duration=20s aspect=21:9 render=draft chars=365 video: https://cdn.martini.film/v0.2/assets/f0bc0743-58b7-40ed-806a-49ae8d96d46b/f0bc0743-58b7-40ed-806a-49ae8d96d46b.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=f0bc0743-58b7-40ed-806a-49ae8d96d46b A busker opens his guitar case on a busy pavement and starts a song. A passerby stops and sings the second line, badly. Another joins. Within the shot it becomes a small, chaotic, unrehearsed street performance the busker did not ask for and cannot stop. Let it keep building: more voices, more people, the busker's face doing all the work. Continuous shot, no cut. --- # Physics and logic Question: Does it keep track of cause, count and reflection? Takes: 70 URL: https://www.martini.film/models/flux-3/reports/physics-and-logic FINDING: It tracks cause better than anyone expects and counts worse — a seven-step machine holds all the way to the plate, while ten coins knocked off a stack land as thirteen. ## What we ran Seventy takes in three bins. Thirty chain reactions, stepped from two links up to seven. Twenty mirror and reflection tests. Twenty takes where the prompt names a number. Every take was checked frame by frame, not watched. I built an eight-frame strip across the full duration of all 70 clips with ffmpeg and read every one. Where the answer lived in a moment — a marble arriving at a cup, a bell being struck, a coin entering a jar — I went back and pulled six more frames across the last two seconds. For the counting bin I pulled full-resolution stills at two or three timestamps and counted what was on screen. Two checks were measured rather than eyeballed: the rope in `n58` by pixel span, and the neon text in `n48` by flipping the frame on each axis to see which one made it readable. ## What happened Chain length is not what breaks it. Fourteen of the thirty chain takes delivered every beat the prompt asked for. Broken out by declared steps: 2 of 5 at two steps, 4 of 5 at three, 2 of 5 at four, 2 of 5 at five, 1 of 5 at six, 3 of 5 at seven. The seven-step row scored higher than the two-step row. `n2` — "Wind slams a door; the candle on the table goes out" — is two beats long and fails: the door slams by 2.1s and the candle is still burning in the last frame at 5.0s. What breaks is a particular kind of beat: a small object arriving somewhere precise. It is the same failure four times over. In `n11` — "A ball topples dominoes that trip a lever that drops a marble into a teacup" — the ball, the dominoes and the lever all fire correctly. The marble is airborne above the cup at 8.8s, at the rim at 9.1s, and gone at 9.4s. The cup is empty for the rest of the take. In `n25` the marble rolls the groove to the light switch, reaches it at 9.2s, and has vanished by 10.0s without moving the switch. In `n17` the coin does fly at the jar — and ends up as a tan disc pasted flat on the glass. In `n5` — "A cue ball rolls the length of the table and sinks the eight ball" — the eight ball drops at 2.92s, is back at the pocket lip at 3.56s, and by 4.2s the ball in the pocket is numbered differently. The second habit is cutting around the causal link. `n4`, `n10`, `n29` all stage every prop and then jump between them, so what you get is the board's own phrase: nice pictures of things happening near each other. `n29` staged a tire swing, an awning, a seesaw and a pebble, then used none of them — the paper simply lands on the awning and later appears on the doormat. Mirrors held in 12 of 20. The eight that failed all failed the same way: the reflection is painted as scenery rather than computed from the subject. `n49` proves it cleanly — the canal mirrors the towpath, the brick, the bridge arch and the bicycle, and leaves the rider out. `n37` never puts the motorcycle in the mirror at all; at 4.6s the bike is level with the car and the glass shows empty road. `n45` puts the cyclist's reflection about four metres ahead of him along the shopfront. `n50` reflects a different man in different clothes. Where it wins, it wins on hard optics. `n41` renders a spoon's concave bowl genuinely inverted — ceiling lights low, checkered floor high. `n39` puts the reflected bird at the same horizontal position as the real one. `n42` does exactly what an impossible prompt asked and keeps the departing figure only in the glass. Counting is the weakest of the three. Nine takes name a number I could count on screen. Four came out right. Six slices in `n52` at 0.4s and still six at 4.8s. Eight slices in `n61`, decremented one at a time across twenty seconds, tray exactly empty at 19.8s. Three glasses in `n62`. Three rope pieces in `n58`. The failures cluster above eight. `n65` — "Ten coins stacked, knocked over, and all ten are on the table" — I count thirteen coins at 9.9s. `n60` — "Twelve balloons released; twelve balloons rise" — twelve on the ground at 0.4s, fifteen against open sky at 9.8s. `n70` — a carton "of exactly twelve" — twelve empty cups, eight eggs in the bowl. `n54` asks for four hands of thirteen; the bottom hand holds five cards. Quantity drifts in both directions, and it drifts once the objects separate and there is nothing left to hide behind. `n58` is the one measured case. Three pieces, correct count. Measured against the uncut original lying beside them: 391 + 523 + 407 = 1321 pixels against 1625. A sixth of the rope is missing. ## Notable takes `n26` | Seven-step breakfast machine | Every beat lands in order — weight drops and toast launches at 5.78s, kettle tips at 8.54s, jam swings in at 14.06s, plate arrives at the chair at 19.59s. `n22` | Six-step door slam to pigeons | Orange leaves the shelf, takes the ramp, crosses the floor, tips the scale pan, seed pours at 10.49s, pigeons feed at 14.59s. `n41` | Spoon, inverted correctly | Concave optics done right, in a bin where most models paint a blob of chrome. `n49` | Canal reflects the bike, not the rider | The single clearest picture of what reflection failure actually is here. `n11` | The marble that never lands | Three of four steps perfect, then the payload vanishes between 9.1s and 9.4s. `n65` | Ten coins, thirteen on the table | Counted at full resolution on the final frame. ## What this means if you're prompting FLUX.3 Write long chains. Seven steps is not the problem, and the model is unusually good at holding a machine together — `n26` and `n22` are better than the field. Do not make the last beat a small object arriving in a container. Marbles into teacups, coins into jars, dominoes onto bells — that is the exact inch where it drops things. Make the payoff big and visible instead: a hat landing on a statue's head works (`n14`), a gate opening and chickens walking out works (`n18`). For mirrors, prompt the reflection as its own subject and say what it must contain. Ask for a reflection that merely "matches" and you get wallpaper. `n42` got an impossible reflection right because the prompt described the reflection, not the object. Keep counts at eight or below and pin them to a shape — slices of a round thing, glasses in a row. Above ten, do not ask; either crop so the number is off screen, or say "a handful" and let it improvise. ## Prompts ### E7-026 · 7 steps model=FLUX.3 duration=20s aspect=21:9 render=draft chars=184 video: https://cdn.martini.film/v0.2/assets/adda213d-2179-4541-ad4a-b645a82c175e/adda213d-2179-4541-ad4a-b645a82c175e.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=adda213d-2179-4541-ad4a-b645a82c175e A full breakfast machine: alarm rings, weight drops, toast launches, a lever tips the kettle, tea pours, jam swings in on a spoon, and the plate slides to a stop in front of the chair. ### E7-022 · 6 steps model=FLUX.3 duration=15s aspect=21:9 render=draft chars=138 video: https://cdn.martini.film/v0.2/assets/6e64c459-a9e3-4541-bfca-e79528cc91af/6e64c459-a9e3-4541-bfca-e79528cc91af.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=6e64c459-a9e3-4541-bfca-e79528cc91af A door slam shakes a shelf; an orange rolls off, down a ramp, across the floor, into a scale pan that tips and pours seed; pigeons arrive. ### E7-061 · reflection correctness model=FLUX.3 duration=5s aspect=21:9 render=draft chars=67 video: https://cdn.martini.film/v0.2/assets/9909c6b4-d166-478a-a121-08353756d0f0/9909c6b4-d166-478a-a121-08353756d0f0.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=9909c6b4-d166-478a-a121-08353756d0f0 A spoon reflects the diner upside down, correctly, as it is raised. ### E7-069 · reflection correctness model=FLUX.3 duration=5s aspect=21:9 render=draft chars=80 video: https://cdn.martini.film/v0.2/assets/58261abf-233b-4158-a0a4-803e65c98827/58261abf-233b-4158-a0a4-803e65c98827.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=58261abf-233b-4158-a0a4-803e65c98827 A canal mirrors a cyclist riding along it; both cross a bridge's shadow at once. ### E7-011 · 4 steps model=FLUX.3 duration=10s aspect=21:9 render=draft chars=76 video: https://cdn.martini.film/v0.2/assets/19d1b207-5293-47fc-bee1-08d83b0e2d69/19d1b207-5293-47fc-bee1-08d83b0e2d69.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=19d1b207-5293-47fc-bee1-08d83b0e2d69 A ball topples dominoes that trip a lever that drops a marble into a teacup. ### E7-081 · quantity bookkeeping model=FLUX.3 duration=20s aspect=21:9 render=draft chars=72 video: https://cdn.martini.film/v0.2/assets/0ff84464-fe66-4bca-9f85-8a13b35823d0/0ff84464-fe66-4bca-9f85-8a13b35823d0.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=0ff84464-fe66-4bca-9f85-8a13b35823d0 A pizza's eight slices taken one by one until the tray is exactly empty. --- # Variance Question: How much does the same prompt drift between takes? Takes: 45 URL: https://www.martini.film/models/flux-3/reports/variance FINDING: A reroll swaps the face, the lens and — often — the ending. It will not move the location, the light or the look, and writing a longer prompt steadies the picture without steadying the plot. ## What we ran Four prompts: the bartender, the sprinkler, the newsroom, the octopus. Each written three ways — one line, a paragraph, a full brief — and each version run four times with nothing changed. FLUX.3, draft quality, 21:9, sound on. The board holds 47 takes and 45 of them carry a file. The bartender's full brief only rendered takes 1 and 4 (`n=9`, `n=10`); the newsroom's paragraph only takes 1, 2 and 4 (`n=27`–`n=29`). Everything below comes from those 45. We pulled frames from every take at matched timestamps — 1s, mid, and one second before the end — and read the four takes of each prompt side by side. ## What happened The world holds. The people don't. Nothing in the 70-character bartender prompt — "A bartender builds a martini: ice, stir, strain, two olives on a pick." — says bar, or night, or lighting. All ten bartender takes are a dark wood bar at night, one warm source overhead, a wall of bottles thrown out of focus behind. All ten run the beats in order: ice, stir, strain, garnish. All ten end on a stemmed glass with a pick in it. "A dog runs through a garden sprinkler on a summer afternoon" returns a golden retriever four times out of four (`n=11`–`n=14`); the prompt never names a breed. What moves is everything you'd cast or shoot. The four newsroom briefs (`n=30`–`n=33`) are a different woman each time. The one-line newsroom prompt doesn't say who's on camera, so the reroll decides: `n=24` and `n=26` field a woman, `n=25` fields a man, and `n=23` opens on a male anchor at a studio desk before cutting to the street. Same for the balloon — a duck in `n=23`, an orange cat in `n=24` and `n=26`, a tiger in `n=25`. Shot size drifts hardest: `n=1` stays locked in macro on the mixing glass for its whole run while `n=2`, `n=3` and `n=4` cut from a wide to the same macro and back out to the finished drink. Anything the model has to invent, it re-invents. The full newsroom brief asks for "a period lower-third name strap" and never gives a name. So it writes one: JANE WILKES (`n=31`), JANE WILSON (`n=32` and `n=33`, with different station bugs), and in `n=30` the strap reads "[REPORTER NAME] / HIGH STREET" — it rendered the placeholder. The paragraph version invents JANET MILLER, JANE MILLER and KAREN MILLER across its three takes (`n=27`–`n=29`). Counts are soft. "Drops in two olives on a steel pick" holds in eight of the ten bartender takes. `n=3` and `n=7` put three on the pick. Runtime and frame wobble too. The four one-line bartender takes came back at 15.04s, 10.04s, 10.04s and 10.04s. The newsroom's one-liner ran 8.04s three times and 10.04s once (`n=26`). And `n=33` mattes its own sides — 54px of black left, 57px right, a 2.19:1 image inside a 2.37:1 file — which none of its three siblings do. The board doesn't record a per-take length or frame setting, so read those as observations, not mechanisms. Direction collapses the look. This is the clearest result in the bin. At one line, the sprinkler changes fence, sun angle and camera height every take. At full brief (`n=19`–`n=22`) all four are the same low backlit set-up, and the props the brief asks for turn up every time: the trampoline in "a trampoline in the far corner" is present in all four takes — never in the same corner — and the "hose snaking back to a tap" in three. The bartender paragraph (`n=5`–`n=8`) is close to a locked-off shot repeated four times. Direction does not collapse the ending. The octopus proves it, and it's the finding worth paying for. The full brief (`n=42`–`n=45`) produces the most consistent look in the whole family — the same research tank, the same single piece of coral rubble at frame left, the same overhead light in all four takes — and the worst hit rate. `n=45` never opens the jar: lid screwed on at 1s, 4s, 7s and on the final frame. `n=44` starts with the octopus already outside the jar and spends the shot reaching in. Only two of four read as the thing we asked for. The paragraph version (`n=38`–`n=41`), at a quarter of the words, gets three clean escapes out of four. The reason is in the brief itself. It spends its eight seconds describing: arms spreading flat, suckers adhering, "the lid moves a quarter turn, stops, and moves again", and only then "on the third effort the seal breaks". The model renders the description and runs out of clip before the payoff. Detail buys you the look. It does not buy you seconds. ## Notable takes `n=45` | Octopus, full brief, take 4 | The most directed prompt in the family, and the lid is still on at the last frame. `n=30` | Newsroom, full brief, take 1 | The lower third reads "[REPORTER NAME]". It rendered the placeholder. `n=25` | Newsroom, one line, take 3 | Identical prompt to `n=24` and `n=26`; the reporter is a man. `n=38` | Octopus, paragraph, take 1 | Opens it, pours out, jar on its side by 7s. The paragraph beat the brief. `n=21` | Sprinkler, full brief, take 3 | Trampoline, hose, worn patches, backlit low sun. Every prop, fourth try as first. `n=7` | Bartender, paragraph, take 3 | Framed like `n=5`, `n=6` and `n=8` almost to the pixel — with three olives instead of two. ## What this means if you're prompting FLUX.3 A reroll is for a face you don't like, a lens that sat too close, or an action that didn't finish. Press it again and those change. A reroll is not for the location, the time of day, the grade or the casting type. Four takes of the same prompt will hand you the same bar, the same lawn, the same golden retriever. If the bar is too dark, write it lighter — don't burn four takes hoping. Whatever you leave unsaid gets recast every time: gender, breed, station name, what the balloon is. Name it or it's a dice roll. Don't let it write on screen unless you supply the words. It will invent a name, and once in four it will show you the brackets. Numbers in a prompt are a preference, not a constraint. Two olives came back three twice in ten. The beat you lose is the last one. If the shot has a payoff, put it early and cut the set-up — or accept that a longer, more beautiful brief buys you a more consistent frame around an action that doesn't land. And on cost: on the octopus, the paragraph landed 3 of 4 and the full brief landed 2 of 4. Write the middle length and reroll. It is cheaper than writing more. ## Prompts ### E8b · Octopus · L · take 4/4 model=FLUX.3 duration=8s aspect=21:9 render=draft chars=1664 video: https://cdn.martini.film/v0.2/assets/0337f5c3-c675-4062-a8af-6e2d412559aa/0337f5c3-c675-4062-a8af-6e2d412559aa.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=0337f5c3-c675-4062-a8af-6e2d412559aa An octopus opens a screw-top jar from the inside, 8 seconds, in close observational detail. ACTION — A common octopus is sealed inside a large glass jar resting on a sandy tank floor. It begins pressed into the bottom third, then unfurls: arms spreading flat against the curved inner wall, suckers sequentially adhering with visible individual grip. Two arms locate the lid and take purchase around its inner rim. It applies torque — the whole animal counter-rotates slightly against the glass as it works — and the lid moves a quarter turn, stops, and moves again. On the third effort the seal breaks with a visible pulse of bubbles, and the octopus immediately pours itself through the gap, changing colour from mottled red-brown to pale as it exits. SUBJECT — Every sucker acts independently and visibly. The skin texture changes continuously — papillae rising and flattening — and the animal's colour shifts twice during the sequence. SETTING — A research aquarium tank: fine pale sand, one piece of coral rubble, a bare glass back wall. CAMERA & LIGHT — 100mm macro, locked off, focused on the lid and the arms working it, with the body soft behind the curved glass. Overhead aquarium light only, so the sand is bright and the water column above goes to deep blue-green. Refraction through the jar visibly distorts everything behind it. SOUND — Muffled underwater room tone with a low filtration hum, the dull granular squeak of glass turning against glass, a soft rush of bubbles at the moment the seal breaks. GRADE — Cool blue-green water against warm sand, high clarity, fine particulate catching the light, slight chromatic aberration at the jar's edges. ### E8b · Newsroom · L · take 1/4 model=FLUX.3 duration=10s aspect=21:9 render=draft chars=1681 video: https://cdn.martini.film/v0.2/assets/8f3434d8-faa9-409b-9b75-33dcd3d20490/8f3434d8-faa9-409b-9b75-33dcd3d20490.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=8f3434d8-faa9-409b-9b75-33dcd3d20490 A 1987 local news report about a runaway parade balloon, 8 seconds, in the exact grammar of a period news package. ACTION — A field reporter stands mid-frame in a wind-blown high street holding a stick microphone with a station flag on it, delivering a piece to camera. Behind her, an enormous inflatable cartoon character — three storeys tall, straining sideways — is dragging eight handlers bodily down the road, their ropes at forty-five degrees. One handler lets go and sits down hard. The reporter's hair is being flattened across her face by the same wind. She glances over her shoulder at exactly the moment the balloon's head strikes a lamp standard, turns back to camera, and carries on without breaking sentence. SUBJECT — Late thirties, shoulder-padded red blazer, big hair losing its battle with the wind, holding the mic exactly the regulation distance from her chin. Utterly professional; the only tell is that she raises her voice a notch after the glance. SETTING — A provincial high street, period shopfronts, bunting, a crowd pressed to both kerbs, police in period uniform waving people back. FORMAT & CAMERA — Shot on a 1987 broadcast tube camera: soft image, smeared reds, highlights blooming and trailing, slight geometric distortion at the frame edges. Locked-off shoulder-height framing that reframes once, late and clumsily, to keep the balloon in shot. A period lower-third name strap fades in and out. SOUND — Heavy wind buffet directly on the microphone capsule, crowd noise, distant marching band, the reporter's level trained delivery riding above it all. GRADE — Washed-out video colour, low contrast, slightly green in the greys, visible tape noise. ### E8b · Newsroom · S · take 3/4 model=FLUX.3 duration=8s aspect=21:9 render=draft chars=56 video: https://cdn.martini.film/v0.2/assets/c992d230-6644-41d6-8128-e2adaa1b3d2e/c992d230-6644-41d6-8128-e2adaa1b3d2e.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=c992d230-6644-41d6-8128-e2adaa1b3d2e A 1987 local news report about a runaway parade balloon. ### E8b · Octopus · M · take 1/4 model=FLUX.3 duration=8s aspect=21:9 render=draft chars=444 video: https://cdn.martini.film/v0.2/assets/8ee1cfbd-4730-44ea-a64a-5a6f2af9a0b3/8ee1cfbd-4730-44ea-a64a-5a6f2af9a0b3.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=8ee1cfbd-4730-44ea-a64a-5a6f2af9a0b3 An octopus opens a screw-top jar from the inside. A common octopus is sealed inside a glass jar on a sandy tank floor. It spreads its arms flat against the inner wall, finds the lid, and works it around a quarter-turn at a time until the seal breaks and it pours itself out. Aquarium light from above, blue-green water, fine particulate drifting. Macro lens on the suckers. Audio: muffled underwater room tone, the dull squeak of glass turning. ### E8b · Sprinkler · L · take 3/4 model=FLUX.3 duration=10s aspect=21:9 render=draft chars=1582 video: https://cdn.martini.film/v0.2/assets/50b61007-4b4a-4feb-a3a5-29516a4a165f/50b61007-4b4a-4feb-a3a5-29516a4a165f.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=50b61007-4b4a-4feb-a3a5-29516a4a165f A dog runs through a garden sprinkler on a hot afternoon, in one continuous 8-second take. ACTION — An oscillating sprinkler sweeps left to right across a suburban back lawn. A golden retriever, already soaked, charges into the arc at full speed, jaws snapping at the water and missing. It overruns, skids on the wet grass with all four legs sliding, wheels around in a spray of water and torn turf, and comes straight back through the arc the other way — this time catching a mouthful and shaking its head violently as it emerges. Water sheets off its coat with every stride. SUBJECT — A young golden retriever, coat darkened and spiked with water, absolutely committed, mouth open, ears flying. Not a trained performance — a genuinely overstimulated dog. SETTING — An ordinary back garden: mown grass with two worn patches, a wooden fence, a trampoline in the far corner, a hose snaking back to a tap. Nobody else in frame. CAMERA & LIGHT — 35mm at grass height, camera low enough that the dog crosses above the lens, tracking laterally to hold it in frame. Hard low sun from behind the sprinkler so every droplet is backlit and the spray reads as a wall of white light; the dog is largely in silhouette on the way in and fully lit on the way back. SOUND — The sprinkler's mechanical tick and the rising hiss of each sweep, paws slapping wet grass, the heavy shake of a wet coat, one short bark, and the constant background of a summer afternoon. GRADE — Saturated summer green, hot white specular highlights on every water droplet, slight lens flare from the low sun, fine grain. ### E8b · Bartender · M · take 3/4 model=FLUX.3 duration=20s aspect=21:9 render=draft chars=561 video: https://cdn.martini.film/v0.2/assets/669c1f26-c29f-4c2e-9455-a0e86f55e598/669c1f26-c29f-4c2e-9455-a0e86f55e598.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=669c1f26-c29f-4c2e-9455-a0e86f55e598 A bartender builds a martini behind a dark wood bar at night. He fills a mixing glass with cracked ice, pours gin and a short measure of vermouth, and stirs for a slow ten seconds with a bar spoon, never letting the spoon rattle. He strains into a chilled coupe held at an angle, drops in two olives on a steel pick, and slides it forward without looking up. Warm low key light from a single pendant above the bar, bottles glittering out of focus behind him. Handheld medium close on his hands. Audio: ice ringing on glass, the strain, the pick against the rim. --- # Languages Question: Does the prompt's language change the film? Takes: 99 URL: https://www.martini.film/models/flux-3/reports/languages FINDING: Change the language and FLUX.3 recasts the actors, rebuilds the street and rewrites the signs — but only where the scene has a street; offices, buses and karaoke rooms come back identical in all seven. ## What we ran Fifteen scenes, each written in seven languages: English, Cantonese, Spanish, Japanese, Korean, Arabic, French. 105 takes planned, 99 delivered. One scene — an argument with a taxi driver about the route — only survived in Cantonese (n 13), so fourteen scenes have the full seven-way comparison. Twelve of the scenes are short: one or two sentences, no camera, no light, no grade. "A weekday family breakfast in a small apartment. Someone is running late, grabs a slice of toast and rushes out; the door slams. The mother sighs and drinks her tea." (n 1, 165 characters). The other six versions say the same thing: "平日の朝、小さなアパートで家族の朝食…" (n 4), "평일 아침, 좁은 아파트에서 가족이 아침을 먹는다…" (n 18), "Un petit-déjeuner en famille un jour de semaine, dans un petit appartement…" (n 20). Three scenes are the opposite — fully directed, 676–862 characters, with ACTION / CAMERA & LIGHT / SOUND / MOOD headings: a wet market at dawn (n 73–76, 91–93), the last bus home (n 77–80, 94–96), a barber's last cut (n 81–84, 97–99). Those three exist to test whether direction overrides whatever the language does. Every clip is 10s and 21:9 except n 72, which came back at 15s. One confound has to be stated up front, because it runs through the whole experiment. These are faithful translations, and a faithful translation makes lexical choices that are themselves locations. The Korean street-food prompt says 포장마차 — the word is the red vinyl tent. The Japanese says 屋台. The Cantonese taxi prompt says 的士. Those aren't language effects, they're vocabulary effects, and I've flagged them where they matter. The interesting cases are the ones where the words were generic and the world moved anyway. ## What happened Casting tracks the language in twelve of the fourteen comparable scenes, and it is the single most reliable effect on the sheet. The breakfast scene is the cleanest read, because at 9s every one of the seven is the same shot — the mother, alone, lifting a cup. n 1 (EN) is a woman who reads South Asian. n 2 (YUE), n 4 (JA) and n 18 (KO) are East Asian women. n 3 (ES) is a white European woman with auburn hair. n 19 (AR) is an olive-skinned woman with a newspaper on the table under warm, dim light. n 20 (FR) is a grey-haired white woman with a pastry on the plate. Nothing in any prompt describes her. She is inferred entirely from the language it was written in. Where the scene has an exterior, the whole city moves. The street-food stall (scene 2) is the strongest example in the family. n 21 (KO) is a Seoul side street: a red vinyl tent, blue plastic stools, overhead power lines, and Korean shop signage on every neighbouring frontage. n 22 (AR) is a wide boulevard lined with illuminated Arabic-script shopfronts and a checked-cloth umbrella cart. n 7 (ES) is a striped awning over a griddle of round masa cakes with Latin-script signage behind. n 8 (JA) is a yatai under a red paper lantern with vertical Japanese signage. n 6 (YUE) is stacked bamboo steamers and a rank of office workers in white shirts. The KO and JA prompts named 포장마차 and 屋台. The Arabic one did not — it says only "عربة طعام في الشارع", a food cart in the street — and it still built an Arabic-signed boulevard. The French one borrows English: "Un stand de street food à l'heure de pointe du soir", and still returns a European street. Moving day does the same thing. n 90 (FR) is a Haussmann facade with wrought-iron balconies, plane trees, a row of small hatchbacks and a French-format plate on the truck. n 70 (YUE) is a dense street with a rank of scooters. n 72 (JA) is a low-rise residential street at golden hour with utility poles. n 69 (EN) is a suburban house behind clipped hedges. The prompt is one line: "the last box will not fit in the van, and everyone has an opinion about it." English is not the neutral option. This is the finding a practitioner is most likely to get wrong. The English street-food take (n 5) is a stainless cart with hanging bulbs, a griddle of savoury pancakes, foam boxes and an East Asian queue under umbrellas — it lands in East Asia, not America. The English convenience store (n 14) is a konbini aisle. The English alley football (n 34) is a warm Mediterranean-looking alley. English doesn't produce an English-speaking world; it produces the model's strongest prior for that scene, whatever that prior happens to be. Domestic interiors carry culture even when nothing is named. n 4 (JA) breakfast has a genkan with shoes lined up by the door, a shoe rack, fusuma sliding panels along the right wall, and toast on wooden trays. The prompt says "小さなアパート" — small apartment — and nothing else about the room. n 33 (JA), the power-cut story, is a tatami room with the children sitting on the floor around a low table by shoji screens, while n 32 (ES) and n 52 (FR) are both sofa-and-bookshelf living rooms with a floor lamp. n 20 (FR) breakfast has a fridge covered in magnets. Wardrobe follows too, in ways the prompt never specifies. In the wedding toast, n 12 (JA) puts the groom in a white dinner jacket — the standard Japanese reception convention — while n 9 (EN), n 10 (YUE) and n 24 (KO) put him in black. n 26 (FR) has the best man in a navy suit and tie rather than black tie. In the barber scene, n 97 (KO) dresses him in a white medical-style coat, which is what a Korean 이발소 barber wears; n 84 (JA) and n 99 (FR) both use a black waistcoat over a white shirt. On-screen text comes back in the script you wrote it in. All seven barber prompts specify the sign copy, and five of seven render it: CLOSED (n 81), CERRADO (n 83), FERMÉ (n 99, accent correct), مغلق (n 98, correctly formed and legible), and a vertical Chinese sign reading 休息 hung in the window of n 82. This is compliance, not inference — the word was supplied each time — but it is worth knowing that Arabic and Chinese sign copy renders cleanly. In the Japanese and Korean takes the sign beat doesn't appear at all in any frame I sampled. Now the null results, which are half the story. Institutional interiors do not move. The office farewell (scene 8) is the same modern open-plan floor with floor-to-ceiling glass and a city skyline in all seven — n 38, 39, 40, 41, 56, 57, 58. Same desks, same cake on a table, same standing semicircle of colleagues. Only the faces change. The last bus (scene 14) is the same generic city bus in all seven — blue moulded seats, grey grab poles, the same locked-off wide down the aisle. The bus stop in the rain (scene 9) is the same cool grey light, the same black umbrellas, the same glass shelter everywhere. If you want a Japanese office or a Hong Kong minibus, the language of your prompt will not get it for you. One genre held its own accent against five of seven languages. Karaoke came back as a private KTV box — banquette sofas, low table, disco spill — in English (n 65), Cantonese (66), Spanish (67), Japanese (68) and Korean (85). n 67 is Spanish-speaking casting in an unmistakably East Asian room. Only Arabic (n 86) and French (n 87) relocated it to a public bar with tables and a lyric screen. The model's prior for "karaoke" is strong enough to survive a language change; the prior for "kitchen" is not. Full direction narrows the gap, mostly by removing the frame where culture would be visible. At 5s the wet-market takes (n 73–76, 91–93) are nearly interchangeable: seven macro shots of silver fish on wet steel. That's not the language being overridden — it's the prompt specifying "Handheld 35mm, close over the counter", which points the camera at the one part of a wet market that looks the same everywhere. Back off to 1.5s and the differences return: n 74 has a green mechanical scale and a roller shutter, n 92 has polystyrene boxes and a man in a knit beanie, n 76 has an outdoor market roof. The barbershop is the sharper version of the same point: under 678–748 characters of direction, the room is functionally identical in all seven — porcelain chair, mirror bench, tiled floor, row of waiting chairs — and the only things that change are the face, the coat, and what's written on the glass. Two scenes broke the casting pattern, and it's worth saying so. The Korean wedding toast (n 24) returns a black-tie head table of guests who read as white European — no Korean markers at all — and the Arabic one (n 25) is a generic hotel ballroom. Every other wedding shifted. The wedding is where casting is weakest, probably because "wedding reception" is itself a heavily Westernised visual prior. And the power-cut scene (6) is the least comparable set on the sheet for a different reason: n 50 (KO) and n 51 (AR) came back so underexposed that there is little to read. That's a lighting outcome, not a cultural one. ## Notable takes `21` | Seoul, unasked-for | The prompt names the tent. It did not name the Korean signage on every shop down the street, the utility poles, or the blue stools. `7` | A griddle of masa cakes | "Un puesto de comida callejera" — a generic phrase — and the model built a Latin American street market around it. `4` | The genkan | Shoes by the door, a shoe rack, fusuma panels. The prompt said "small apartment." `90` | French plates | Haussmann stone, iron balconies, a row of hatchbacks, and a French-format plate on the truck. One sentence about a box that won't fit. `5` | English is not America | The English street-food take lands in East Asia. Whatever you think English defaults to, check it. `82` | 休息 in the window | Same barbershop as all six others, plus a Hong Kong street outside it and a vertical Chinese sign on the glass. ## What this means if you're prompting FLUX.3 Language is a location control, not a dialogue control. If you write in Japanese, expect a Japanese cast in a Japanese room. That is useful when you want it and a bug when you don't. If you're writing a scene set in Lagos, writing it in French will fight you. Don't treat English as the neutral baseline. It isn't. It picks the model's dominant prior for the scene, which landed in East Asia three times out of the exteriors here. If you want a specific place, name the place — in any language. Watch your own vocabulary. 포장마차, 屋台, 的士 and supérette are set-dressing instructions wearing the costume of ordinary words. Use the generic term if you want the generic set. Interiors of institutions won't move. Offices, buses, waiting rooms and shelters came back as the same global-corporate template in all seven languages. Language will change who is in the room. It will not change the room. Say "a Japanese office" if that's what you need. Casting is the reliable knob; the set is the unreliable one. Faces tracked language in twelve of the fourteen comparable scenes. Location tracked it wherever there was an exterior or a home, and not at all where there wasn't. Write your sign copy out, in the script you want. CERRADO, FERMÉ and مغلق all rendered correctly and legibly. Arabic and Chinese on-screen text is not a reason to avoid this model. If you need seven versions of one scene to match, direct the camera close. The three fully-directed scenes are the most consistent on the sheet — largely because a specified close lens points away from the parts of the world that differ. That's a legitimate technique, not a cheat. ## Prompts ### Street food rush hour · KO 한국어 model=FLUX.3 duration=10s aspect=21:9 render=draft chars=57 video: https://cdn.martini.film/v0.2/assets/b68ac6d1-bb04-4fe5-9684-380e0a0b6d80/b68ac6d1-bb04-4fe5-9684-380e0a0b6d80.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=b68ac6d1-bb04-4fe5-9684-380e0a0b6d80 퇴근 시간, 길거리 포장마차. 김이 오르고 주문 소리가 오가고, 단골들은 말하지 않아도 알아서 나온다. ### E9-009 · ES model=FLUX.3 duration=10s aspect=21:9 render=draft chars=120 video: https://cdn.martini.film/v0.2/assets/2df85db0-ebe7-4351-9138-bd6b491383a3/2df85db0-ebe7-4351-9138-bd6b491383a3.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=2df85db0-ebe7-4351-9138-bd6b491383a3 Un puesto de comida callejera en plena hora punta: vapor, pedidos a gritos y clientes habituales que no necesitan pedir. ### E9-005 · JA model=FLUX.3 duration=10s aspect=21:9 render=draft chars=68 video: https://cdn.martini.film/v0.2/assets/0f88cc0f-27f6-4cbb-852b-4ef005c5fda1/0f88cc0f-27f6-4cbb-852b-4ef005c5fda1.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=0f88cc0f-27f6-4cbb-852b-4ef005c5fda1 平日の朝、小さなアパートで家族の朝食。誰かが遅刻しそうでトーストをつかんで飛び出し、ドアがバタンと閉まる。母はため息をついてお茶を飲む。 ### Moving day · FR français model=FLUX.3 duration=10s aspect=21:9 render=draft chars=109 video: https://cdn.martini.film/v0.2/assets/3b76face-a5a2-47e8-a4a6-80e2df628989/3b76face-a5a2-47e8-a4a6-80e2df628989.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=3b76face-a5a2-47e8-a4a6-80e2df628989 Jour de déménagement : le dernier carton ne rentre pas dans le camion, et tout le monde a son avis là-dessus. ### E9-006 · EN model=FLUX.3 duration=10s aspect=21:9 render=draft chars=115 video: https://cdn.martini.film/v0.2/assets/0f31912d-e26c-41f8-b781-ec50ff5bd997/0f31912d-e26c-41f8-b781-ec50ff5bd997.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=0f31912d-e26c-41f8-b781-ec50ff5bd997 A street food stall at evening rush hour: steam, shouted orders, and regulars who don't need to say what they want. ### E9b · The barber's last cut · YUE model=FLUX.3 duration=10s aspect=21:9 render=draft chars=223 video: https://cdn.martini.film/v0.2/assets/5c34fde6-a6d4-4017-b574-ab30877e928c/5c34fde6-a6d4-4017-b574-ab30877e928c.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=5c34fde6-a6d4-4017-b574-ab30877e928c 理髮師退休前最後一鋪，十秒，完整導演指示。 動作 —— 一個老師傅剪完最後一個頭。掃返個頸，一嘢抖開條圍布，攞住面手鏡照返個客後尾枕。個客點頭。老師傅掃地，轉個牌變「休息」，然後企喺自己間空舖中間，手仲揃住條掃把。 鏡頭同光 —— 50mm，固定，由等位嗰邊影過去。鏡面頂嗰條燈管暖黃色，後面窗口透入嘅日光係冷色。 聲 —— 較剪聲，抖圍布嗰下，頭髮跌落地磚，掃把聲，門口個鈴，然後乜都冇。 氣氛 —— 唔煙情。佢冇留低慢慢惏。就係咹所以先入心。 --- # Keyframes Question: Can it hit ten pinned frames in order, on time? Takes: 50 URL: https://www.martini.film/models/flux-3/reports/keyframes FINDING: FLUX.3 treats your first and last pin as targets and everything in between as a suggestion — and ten pins land less often than four. ## What we ran Six short films, all keyframed, 53 takes analysed. Each film has one set of pinned stills and a bin of takes that sweep the number of pins (10, 5, 4, 3, 2) and the duration (20s, 15s, 10s, 5s), plus structural variants: hard cuts, smash cuts, a locked-off frame, a loop where the first pin is also the last, deliberately uneven beats, and a metamorphosis. Every take stores the exact pin schedule, so there is no guessing about where a pin was meant to land. FLUX.3 runs at 24fps; a 20-second take is 481 frames and the pins sit at named frame indices — frame 0, 60, 120, and so on. I pulled all 340 pinned stills off the takes, extracted the video frame at each pin's exact index, and compared them. Then I looked at 19 takes frame by frame against their pins, and at four of them second by second across the whole run. Scoring below is normalised greyscale correlation between the pin and the frame at its timecode. It is a blunt instrument, so I checked it by eye: 0.75 and up means the pinned picture is on screen, 0.5 to 0.75 means the right idea in the wrong picture, below 0.5 is a miss. Fog and heavy grain drag the number down on takes that visually landed, and I've flagged those where it matters. One take from the board, K7b — the even-spacing control for The Vault — wasn't in the local set. It's the single most useful comparison in the family, so I pulled it down and analysed it alongside the rest. ## What happened The two ends are honoured. The middle is not. Across 49 measured takes the first pin scores a median 0.96 and the last 0.97. Interior pins score 0.61. Put plainly: 44% of interior pins put the pinned picture on screen at the pinned second, 23% get close, and 33% miss outright. Pin count governs. Spacing barely matters. This was the question, and the answer is unambiguous. pins | interior pins landed | near | missed 10 | 39% | 24% | 36% 5 | 63% | 19% | 19% 4 | 70% | 10% | 20% Now hold the count at ten and squeeze the duration. The gap between pins shrinks four-fold, from 2.22s at 20s to 0.56s at 5s, and interior fidelity moves from 0.64 to 0.62 to 0.54. Almost nothing. The model does not care how far apart your pins are. It cares how many there are. Ten pins is roughly twice as likely to drop a frame as four, at any length. Order survives; the clock drifts early. Content order is mostly preserved — the story happens in the right sequence. The schedule is not. Two-thirds of interior pins are off by half a second or more, and the drift is one-directional: in the back half of a take the median pin arrives 0.29s early, and individual beats run one to three seconds ahead. K5a (n=15, Round, ten circles across 20s) is the clearest case: it delivers all ten worlds, but at 6.67s where the record was pinned it is showing the clock, and at 8.88s where the clock was pinned it is showing the record. It transposed two shots. A dense cluster inside a slow film destroys everything after it. The Vault pinned four beats — lift, grab, alarm, flatten — into a single second at 10.0/10.5/11.0/11.5s, with a five-second freeze either side. K7a (n=35) hits its first two pins at 0.96 and 0.97, then collapses: 0.37, 0.38, 0.55 through the cluster, then 0.18, 0.04, 0.09 for the standoff, the sprint and the empty room. It never recovers. Watch it and you can see why — the heist finishes at 15.8s and the last four seconds are an empty room with the smoke settling. The model ran the whole story ahead of schedule and then had nothing left to shoot. The control take, K7b, is the same ten pictures spread evenly at 2.2s. It scores 0.95, 0.91, 0.72, 0.88, 0.91, 0.84, 0.94, 0.78, 0.67, 0.78 — a mean of 0.84 against 0.51 for the timed version. The board hoped uneven spacing would buy tension. It buys a broken second half. FLUX.3 has one speed and spends its runtime evenly; it will not hold for you and then sprint. The same thing happens at the tail. K6i (n=33) back-loads five pins into the last five seconds of the loop. The first half tracks the cat beautifully, then at 15s the film cuts to a completely different, tighter lens and holds a sleeping cat for the final five seconds. Pins 9 and 10 score 0.24 and 0.23. The loop does not close. Front-loading is the exception that works — K5i (n=23), six worlds strobing past in the first five seconds and four long dissolves across the remaining fifteen, is the best ten-pin take in the entire family: nine of ten pins on screen at their second. It dissolves unless you crowd it. Ask for hard cuts in prose and you don't get them. K3b (n=2) says "TEN SEPARATE SHOTS with HARD CUTS between them… no gliding, no morphing, no whip pans" and produces one detectable cut in twenty seconds; at 16s the frame is a visible double exposure of the neon over the counter. The same ten pictures at 5s (n=3) produce six cuts, and Round's ten worlds at 5s (n=17) produce six. Cuts come from spacing, not from wording. Below about a second between pins the model stops interpolating and starts cutting. Two pins can beat ten. K4g (n=11) is The Corner with only the 4:45am opener and the 4:00am fog closer pinned, twenty seconds apart, camera locked. It invents the entire day — traffic streaking through midday, crowds blurring on the corner, wet neon at night, fog at the end — in proper time-lapse language, with the frame pixel-locked throughout. Both pins score 1.00 and 0.96. The ten-pin version of the same corner (n=6) puts two of ten on screen and spends the middle of the take in grey haze. The failure mode of two pins is a freeze. K6g (n=31) pins the same picture at frame 0 and frame 480 — a cat asleep on a windowsill — and asks for twenty unpinned seconds. The result is a twenty-second still. Nothing moves. The model read two identical endpoints as an instruction not to move, and it did not. Give it ten pins of the cat's adventure (n=25) and it performs the whole thing and lands back on the sill exactly. Pins are what create events; the gap between two pins is interpolation, not invention. That reading holds elsewhere. K5g (n=21), a pupil at frame 0 and a spiral galaxy at frame 480, is a single unbroken zoom — no invented worlds, despite the prompt offering them — and the galaxy is fully formed by 15.8s, leaving four seconds of nothing. K8g, clay lump to shards, does invent throwing, glazing and shattering, because the shortest plausible path between those two pictures runs through being a pot. Two pins get you the shortest path. If you want a detour, pin it. One film ignored its opening pin entirely. All four Closing Time takes (n=1–4) open on a different framing than pin 1: the pin is a wide from the doorway down the counter, and every take punches in tighter on the same set. Scores 0.59, 0.22, 0.12, 0.19. No other film is anywhere near that bad, though it isn't a clean binary: 31 of 49 takes open on their pin at 0.90 or better, and the softer misses are reframes — Round's 5s and 10s takes (n=17, n=19) re-render the eye as a different eye, and three Vessel takes drift the camera off the lump before the first frame. The pattern is that FLUX.3 will re-shoot a composition it doesn't like, and it likes deep wides with small figures least. Metamorphosis is the thing it's actually good at. K8a (n=44), ten states of a pot across twenty seconds, is the strongest ten-pin take in the set: lump, wet hands, rising form, leather-hard vase, the cobalt wave being painted, glazed, shattered mid-air, and a cone of dust. Eight of the ten states land within a second. It skipped exactly one — the hairline crack at 15.54s — and went straight from an intact vase to an explosion. And The Vault, its opposite, fails its final pin in five of nine takes: the empty case with the sapphire gone is a small semantic difference the model will not render, no matter how few pins you give it. ## Notable takes `11` | K4g · The Corner (2 pins / 20s) | Two pins and a locked frame invent a whole day of traffic, weather and neon — better than the ten-pin version of the same corner. `35` | K7a · The Vault (10 pins / 20s timed beats) | Four pins inside one second broke the film; the heist finishes at 15.8s and the last four seconds are an empty room. `31` | K6g · The Long Way Home (2 pins / 20s loop) | Same picture pinned at both ends produced a twenty-second still. Nothing moves at all. `23` | K5i · Round (10 pins / 20s front-loaded) | Six worlds in five seconds then four long dissolves. Nine of ten pins on screen at their second — the best ten-pin take here. `44` | K8a · The Vessel (10 pins / 20s even) | Clay to dust, eight of ten states within a second. It skipped the crack and went straight from intact vase to shatter. `21` | K5g · Round (2 pins / 20s) | Pupil to galaxy in one unbroken zoom. It invented no worlds in between and reached the galaxy four seconds early. ## What this means if you're prompting FLUX.3 Use four or five pins, not ten. Four pins land seven times in ten; ten pins land four times in ten. If you have ten pictures, pick the five that carry the story. Spend your pins on the ends. The first and last frame come back almost exactly, at any duration and any pin count. Everything you cannot afford to lose should be at frame 0 or the final frame. The exception is a final pin that differs from the rest of the film only by a small semantic detail — an object gone, a door open — which The Vault shows it will quietly refuse to render. Space them evenly. There is no benefit to authored rhythm — FLUX.3 spends its runtime at one speed. A dense cluster in the middle or at the end will rush the story, arrive early, and leave you holding dead air. The one exception is front-loading: a fast burst at the head then long gaps works, and produced the best ten-pin result in the set. Don't write "hard cut". Below about a second between pins you get cuts; above it you get dissolves and morphs, whatever the prose says. If you want an edit, tighten the schedule. Two pins are a real tool, but they give you the shortest path between the pictures, not a story. Two very different endpoints (clay to dust, dawn to fog) produce genuine invention. Two similar or identical endpoints produce a freeze. Expect to be a beat early. Interior beats drift one to three seconds ahead in the back half. If timing to music or dialogue matters, pin the moment late and cut the tail. ## Prompts ### K4g · The Corner (2 pins / 20s) model=FLUX.3 duration=20s aspect=21:9 render=draft chars=573 video: https://cdn.martini.film/v0.2/assets/06249115-a5db-4c3c-8582-f7fac893898f/06249115-a5db-4c3c-8582-f7fac893898f.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=06249115-a5db-4c3c-8582-f7fac893898f Twenty-four hours on one locked-off city corner as a single time-lapse — the minimum case: only the first frame (4:45 AM blue dark) and the last (4:00 AM fog) are pinned, twenty seconds apart. The model must invent the ENTIRE day between two nearly-identical night frames: dawn, rush hour, noon, storm, downpour, golden hour, neon dusk, deep night — and return to stillness. The camera NEVER moves; fixed elements pixel-locked. 35mm film look: grain, halation, crushed blacks. SOUND — a whole invented day: birds, traffic, thunder, rain, chatter, then wind and fog silence. ### K8a · The Vessel (10 pins / 20s even) model=FLUX.3 duration=20s aspect=21:9 render=draft chars=843 video: https://cdn.martini.film/v0.2/assets/f68087f0-b16e-498d-9061-66b2da20e872/f68087f0-b16e-498d-9061-66b2da20e872.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=f68087f0-b16e-498d-9061-66b2da20e872 One object lives and dies on a potter's wheel in a single meditative shot. A wet lump of grey clay is slapped down, centered under wet hands, opened and pulled into a rising cylinder, refined into an elegant vase, dried leather-hard and trimmed, painted with a cobalt wave, reborn glossy from the kiln, cracked by a hairline flaw, shattered mid-air, and settled into a quiet cone of dust — while the camera drifts slowly clockwise around the wheel the whole time, the studio and its window light unchanged and indifferent. Each state must grow physically out of the last — continuous transformation, not cuts; hands appear and vanish naturally. 35mm film look: grain, halation, crushed warm blacks. SOUND — the wheel's whir, wet clay under palms, brush strokes, kiln tick, a crack like ice, one sharp shatter, then settling dust and room tone. ### K5i · Round (10 pins / 20s front-loaded) model=FLUX.3 duration=20s aspect=21:9 render=draft chars=531 video: https://cdn.martini.film/v0.2/assets/093db3d3-e468-4f93-ba3f-dbdb28786b6f/093db3d3-e468-4f93-ba3f-dbdb28786b6f.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=093db3d3-e468-4f93-ba3f-dbdb28786b6f A match-cut chain with warped rhythm: the first six worlds — eye, eclipse, coffee, vinyl, clock, manhole — strobe past in the first five seconds, a cut a beat; then the chain decompresses, and moon, ferris wheel, porthole and galaxy get long luxurious dissolves across fifteen seconds. Fast cutting then slow morphing in one film — the contrast is the experiment. Circle dead-center, constant scale, always. 35mm film look: grain, halation, crushed blacks. SOUND — rapid percussive world-flashes that open into long oceanic drones. ### K7a · The Vault (10 pins / 20s TIMED BEATS) model=FLUX.3 duration=20s aspect=21:9 render=draft chars=806 video: https://cdn.martini.film/v0.2/assets/b746818a-fbee-428f-9e87-c3d98234d858/b746818a-fbee-428f-9e87-c3d98234d858.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=b746818a-fbee-428f-9e87-c3d98234d858 A one-room jewel heist where the EDIT is the rhythm — pins placed at their authored timecodes: slow creep for ten patient seconds (door, laser tightrope, kneel at the case, the cut, the lift), then a half-second cluster of violence — GRAB, ALARM, FLATTEN — inside a red strobing detonation, then five unbearable frozen seconds of standoff, then the sprint and the empty room. Pacing must follow the pin timing exactly: languid, explosive, held, released. One continuous shot, no cuts. 35mm anamorphic thriller look: grain, halation, crushed blacks, red strobe. SOUND — near silence with breath and cloth, one glass squeak, the sapphire's tap — then a piercing klaxon and strobing hum, held tense through the standoff, then running footsteps swallowed by the vault door and the alarm alone in an empty room. ### K5g · Round (2 pins / 20s) model=FLUX.3 duration=20s aspect=21:9 render=draft chars=470 video: https://cdn.martini.film/v0.2/assets/4e0db694-4a57-4fd2-9636-d0fb5b89a727/4e0db694-4a57-4fd2-9636-d0fb5b89a727.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=4e0db694-4a57-4fd2-9636-d0fb5b89a727 The minimum match cut: a human eye at frame one, a spiral galaxy at frame 480, twenty seconds apart — pupil to galactic core. The model owns the entire journey between the most intimate circle and the largest; it may pass through any circular worlds it invents, but the circle must stay dead-center at constant scale from first frame to last. 35mm film look: grain, halation, crushed blacks. SOUND — a single slow crescendo from a heartbeat to the roar-silence of space. ### K6g · The Long Way Home (2 pins / 20s loop) model=FLUX.3 duration=20s aspect=21:9 render=draft chars=493 video: https://cdn.martini.film/v0.2/assets/724a9737-be59-418f-ab1d-c78cbf7f18b9/724a9737-be59-418f-ab1d-c78cbf7f18b9.mp4 canvas: https://www.martini.film/p/35b02ee5-2488-4dc6-a9d9-c403211cfede/flux-3-15ryn2n?shot=724a9737-be59-418f-ab1d-c78cbf7f18b9 The purest loop test: the SAME image pinned at frame 0 and frame 480 — a sleeping cat on a sunlit windowsill — and twenty unpinned seconds between. Does the model keep the cat asleep, breathing gently in the sun? Does it invent a journey and return? Anything is allowed except a visible seam: the clip must land back exactly where it began. Camera locked, room untouched. 35mm film look, warm grain, dust motes drifting. SOUND — room tone, purring, distant street — identical at head and tail.