Flux 3: Scenes, Not Clips

Nobody's afternoon gets ruined by a mediocre clip. You regenerate it and move on.

What ruins the afternoon is the hour after. Cutting four takes together so the light matches across the seams. Running dialogue through a voice model, then nudging it frame by frame until the mouth stops lying. Hunting foley for a door that closes at six seconds. Extending a shot and discovering the camera drifted a metre to the left and the room quietly rearranged itself while you weren't looking.

So the interesting question about a video model isn't how good the output looks. It's how much work survives after the output arrives.

Flux 3 is Black Forest Labs' answer to that question, and the answer is unusual: sound and picture come out of the same generation. Not audio attached afterward by a second system. The same pass that decides where the character's hand goes also decides what it sounds like when it lands.

What Flux 3 Actually Is

One model, trained jointly across image, video and audio on a unified backbone the lab calls Self-Flow. That architectural detail sounds like a footnote and isn't.

If a model learns audio separately and staples it on at the end, the sound will be plausible and generically so. It knows what a kitchen sounds like. It doesn't know what this kitchen sounds like at the moment this glass tips over. Learn both together and the audio becomes causally tied to what's happening on screen. Footsteps land when feet land. The alarm pulses with the light.

That is the whole bet. Everything else follows from it.

Twenty Seconds Is a Different Unit

Flux 3 generates up to twenty seconds in a single pass, with presets at five, ten, fifteen and twenty, or auto if you'd rather the model read the pacing from your prompt.

The number matters less than what the number holds. Five seconds is a gesture — a look, a turn, a hand reaching. Twenty seconds is a complete beat. Setup and payoff. Someone enters, something changes, the shot resolves. You can put an idea inside twenty seconds in a way you simply cannot inside five.

Which means fewer takes to stitch. Fewer seams to hide. The arithmetic is boring and the effect is not: every seam you don't have to hide is a decision you don't have to make, and creative work dies by a thousand small decisions more often than by one big failure.

Sound That Arrives With the Picture

Native audio in Flux 3 covers ambience, sound effects, music and dialogue — including dialogue in multiple languages, with lip sync, and with characters who can speak different languages inside the same scene.

Think about what that removes for a team shipping the same thirty-second spot into four markets. Normally that's a separate localisation pass with its own vendor, its own timeline, its own round of fixes when the Turkish take runs two seconds long and the mouth stops matching. Generate the dialogue with the picture and the timing problem never opens.

There's a toggle, and you should know when to reach for it. If you're prototyping, leave audio on and get a feel for the whole clip in one go. If a sound designer is going to own the mix anyway, switch it off you'll spend less time and you won't be fighting a soundtrack you're about to throw away.

Four Ways Into the Same Shot

Here's the part most model write-ups skip, and it's the part that decides which entry point you reach for.

There are four Flux 3 video models on each::labs, and they aren't interchangeable. Each one assumes a different amount of certainty about what you want.

Text to Video is the open end. You have a description and no footage. Prompt, duration, aspect ratio, audio on or off. Best when you're exploring — when the point of the generation is to find out what the idea looks like.

Image to Video anchors one thing: the look. You bring a still and the model moves it. This is where a product photo becomes a hero shot, or concept art becomes a teaser, because the composition and the palette are already decided and only motion is in question.

Image to Video Timestamped is the one people overlook, and it's the most interesting of the four. You pass an array of keyframes, each with a timestamp. Frame at two seconds. Frame at four. Frame at six. The model fills the motion between them.

That flips the relationship. You're not describing an outcome and hoping. You're setting marks and asking the model to move between them. If you've ever storyboarded a shot and then watched a generator ignore your third beat entirely, this is the model that fixes it. Product reveals, timed transitions, anything where a client has approved specific frames and you need those exact frames to appear at those exact moments.

Video Continuation is for when the footage already exists. Feed it a clip and tell it what happens next.

Read those four in order and a gradient appears: no anchor, one anchor, several timed anchors, existing footage. The more you already know about the shot, the more the model lets you pin down. That's a rare shape in generative tooling, where the usual choice is between total control and total abdication with nothing in between.

Continue the Shot Without Losing the Room

Continuation deserves its own beat, because it's the feature most likely to save you and the one most likely to disappoint you if you misread it.

What it does well: it carries the scene forward. Composition, style, colour, camera behaviour, and — because audio is generated jointly — the soundtrack across the seam too. The ambience doesn't restart. The room stays the room. Anyone who has extended a shot only to watch the grade shift and the walls move will understand why that's worth something.

What it does not do: know where your story was going. Continuation preserves the room, not the intent. Say what should happen next, plainly, and it will take you there. Stay vague and you'll get twenty seconds of something technically consistent and dramatically inert.

One practical note: it wants clean input. Fast cuts, heavy motion blur and busy overlays all make it harder for the model to hold semantics steady from frame to frame. A stable shot with a clear subject continues far better than a chaotic one.

Where People Actually Use Flux 3

Agencies working to a delivery date, where a failed generation isn't an inconvenience but a missed handoff. Twenty seconds with sound attached means the first cut you show a client is watchable rather than a silent animatic you have to talk over.

Brands and retailers producing across markets, where multilingual dialogue removes an entire vendor from the chain. Post teams folding generation into an existing pipeline, using timestamped keyframes so what comes back matches what was approved rather than an interpretation of it.

Creators shipping volume, where the constraint was never ideas but turnaround.

And product teams building AI-native apps, where the sharpest move is exposing continuation to end users extend this, keep everything else. That's a feature people understand without a tutorial.

Getting Better Results Out of Flux 3

Write sound into the prompt, not around it. Audio comes from the same semantic signal as the picture, so naming the ambience gets you a soundtrack that fits. "Rain on a metal awning, traffic two streets away" does real work. Leaving audio unmentioned doesn't mean silence it means the model guesses.

Describe motion in time, not in adjectives. "Cinematic" tells it almost nothing. "Slow dolly forward, camera settles as she turns" tells it what to do across the clip. Duration is a canvas. Say what happens where on it.

Resist over-specifying. The counterintuitive one. Stacking every instruction you can think of, especially in continuation, can conflict with the dynamics already present in your input and produce artifacts or abrupt scene changes. Say the important things clearly and stop.

Test short, then extend. Validate motion at five or ten seconds before spending a full twenty on something you haven't seen yet.

Using Flux 3 on each::labs

All four models run through the same endpoint. Moving between them is a change of model string and input shape, not a new integration:

bash

curl -X POST \
  -H "Authorization: Bearer $EACHLABS_API_KEY" \
  -H "Content-Type: application/json" \
  --data '{
    "model": "flux-3-text-to-video",
    "version": "0.0.1",
    "input": {
      "prompt": "Slow dolly forward through a rain-wet alley at night, neon reflections on the ground, distant traffic and rain on metal",
      "duration": "15",
      "resolution": "hd",
      "aspect_ratio": "16:9",
      "generate_audio": true
    },
    "webhook_url": ""
  }' \
  https://api.eachlabs.ai/v1/prediction/

Point model at flux-3-image-to-video and add image_urls. At flux-3-image-to-video-timestamped and pass a keyframes array where each entry carries an image_url and a timestamp. At flux-3-video-continuation and pass a video_url. Same auth, same call, same response handling.

The call returns a prediction ID rather than a video, which you poll or receive by webhook. That shape isn't an inconvenience to work around — generation latency varies, and any integration assuming a synchronous response breaks the first time it meets real traffic. Aspect ratios run from 21:9 through 9:21, with auto to match your source.

The Honest Limitations

I don't want this to read like a brochure with the rough edges sanded off, so here they are.

Resolution is the big one. Current outputs sit at 480p and 720p. Native 4K is not part of the public specs and shouldn't be assumed. If your deliverable is a broadcast master, this is not yet the tool for the final render — it's the tool for everything upstream of it.

Twenty seconds is a hard ceiling per generation. Anything longer is a multi-pass workflow where you chain clips and assemble. Better than five-second fragments, still not a scene you can shoot in one call.

It isn't fast. Median runtime across all four models sits around four minutes. Fine for a production queue, wrong for anything interactive. Design your UX around waiting.

Audio is real but not finished. Black Forest Labs has been open that audio makes up a very small share of the training tokens relative to video. In practice that means ambience and effects land well, while intricate musical structure and precisely scripted speech fall short of what a sound designer would sign off on. Treat generated audio as a strong first pass, not a final mix.

Early access shows. Behaviour varies more than a mature system, particularly on edge-case footage, and the published benchmark numbers are the lab's own, measured on a pre-release candidate. Read them as a floor rather than a verdict. Performance characteristics are still moving.

Timestamped is the least documented of the four. Exact resolution ceiling and latency for that specific mode aren't separately confirmed anywhere public. It works, and it's genuinely useful, but test it against your own requirements before you build a pipeline on it.

Wrapping Up

The reason Flux 3 is worth your attention isn't the twenty seconds, and it isn't the audio on its own. It's that the output arrives closer to finished than it used to, and that the four entry points let you decide how much of the shot you're willing to hand over.

You can bring nothing and explore. You can bring a frame. You can bring a set of frames with times attached and get back exactly the beats you planned. You can bring footage and extend it without losing the room.

The clip was never the deliverable. This is the first family of models that seems to have noticed.

You can try Flux 3 on each::labs.

Frequently Asked Questions

What makes Flux 3 different from other video generation approaches?

Where the audio comes from. Generate silent video and attach sound afterward and you get sound that's plausible but unconnected to the specific events on screen. Flux 3 produces both from the same pass, so what you hear is tied to what happens. Whether that matters to you depends on whether your pipeline currently has a sound step in it. If it does, this collapses it.

Which Flux 3 model should I use?

Ask how much of the shot you already know. Only a description, use Text to Video. A still you want moving, Image to Video. Specific frames that must appear at specific moments, Image to Video Timestamped worth learning if you work to approved storyboards. Footage you want extended, Video Continuation. Four models, ordered by how much of the shot is already decided.

Can Flux 3 handle dialogue in languages other than English?

Yes, including different languages from different characters inside one scene, with lip sync. The caveat: pronunciation quality varies by language and needs review, and precisely scripted lines aren't where the audio side is strongest. For a market-localised campaign it removes a dubbing pass. For a dialogue-led narrative piece, budget review time.