FLUX.3 is in Martini.
Black Forest Labs' new video model — twenty-second takes with native audio, ten keyframes pinned to exact timecodes, and a draft mode you can render up. Available today.
Open the FLUX.3 test canvas971 takes, every prompt attached. No sign-in.
Three of the thousand
Vault, 20 Seconds
Four shots, one generation, rendered up from draft
The Wire, Kowloon
Cantonese dialogue, subtitled
Deux Parapluies
French New Wave, jump cuts, freeze frame
Straight out of the model. Nothing was cut, graded, or retouched.
What it does
Longer takes than anything else we run, and the two controls that make them usable.
- 20s
- One generation runs up to twenty seconds at 24fps, with native audio.
- 10
- Keyframes pinned to exact timecodes. First and last frame too.
- Draft
- Rough a take cheaply, then render that same take at 1080p.
- Extend
- Carry a shot on from where it ended.
What it costs
You pay by the second. Drafting costs a fraction of rendering, so the takes you throw away barely cost anything.
| Render | Per second | A 20s take | Extend, /s |
|---|---|---|---|
| DraftRough a take at 720p. Nothing is committed. | $0.050.5 ol | $1.00 | $0.11 |
| 720pFull quality, standard resolution. | $0.151.5 ol | $3.00 | $0.37 |
| 1080pFull quality, full resolution. | $0.262.6 ol | $5.20 | $0.48 |
| EnhanceRe-render a draft you kept. Always comes back 1080p. | $0.282.8 ol | $5.60 | — |
Olives are Martini's credit and 1 ol = $0.10. Nothing is charged per generation or per seat on top.
A worked example
Roll 10, keep 1: 70% cheaper.
Say one 20-second shot takes 10 tries to land. Draft all 10, enhance only the keeper, and it comes to $15.60 instead of $52.00. Roll fewer takes and you save less — the rates themselves don't change.
If you only ever roll once, skip drafting: a draft plus an enhance is $0.33 a second, where going straight to 1080p is $0.26. Drafting wins from the second take on.
Extending a shot works the same way: a 20-second continuation is $2.20 drafted, or $9.60 at 1080p. Extends cost more than a fresh take, because the model has to read the clip you already have before it can write the new one. Draft the join first.
Techniques
FLUX.3 is better at reasoning than at arithmetic, and better at building a world than at cutting one.
13 experiments, 971 takes, on what actually moves the picture and what only looks like it should. Each one is written up in full, with every prompt.
- FormatsCan it hold a format — and what happens when you contradict it?
- On-screen textCan it write — title cards, dates, scorebugs, lower thirds?
- Prompt styleWhich way of writing a prompt wins?
- Multilingual scenesHow many languages can share one scene?
- Prompt lengthDoes a longer prompt make a better film?
- Finished filmsCan it make something that reads as an actual film?
- ErasHow far back does its sense of period go, and what happens at the seams?
- Sound and dialogueCan sound drive the picture rather than decorate it?
- DurationDoes twenty seconds buy a story?
- Physics and logicDoes it keep track of cause, count and reflection?
- VarianceHow much does the same prompt drift between takes?
- LanguagesDoes the prompt's language change the film?
- KeyframesCan it hit ten pinned frames in order, on time?
Read all 13 reports
971 takes, 184 minutes of video, and every prompt that produced them.
FLUX.3 is live in Martini today, and every take from this write-up is on one canvas.
Open the FLUX.3 test canvas