Duration

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.

1A 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.

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.

1Pre-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.

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.

1A 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.

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.

1A man finishes telling a story at a dinner table and stops talking. The camera does not cut.

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.

1A 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.

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.

1A 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.

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.