FLUX 3

Black Forest Labs' first video model, and the one place in this catalog where a clip is written rather than specified. The prompt is interpreted and expanded before generation, so full sentences work and keyword lists do not: describe the scene, the action, the camera, and the sound you want in the order you would say them out loud. Every other clip model here rewards terse motion-verb syntax. This one rewards prose.

The sound is the other half of the same prompt. Ambient noise, effects and spoken lines come out of the same sentence as the picture, in sync, and can be switched off for a silent take. Dialogue in quotation marks gets mouth movement to match.

Its window is five to twenty seconds, or you can leave the length to the model. Twenty seconds at 1080p is the longest high-resolution take on offer here — Seedance 2.5 runs to thirty but stops at 720p, and Veo caps out at eight. It is also the most expensive: 71 credits for five seconds at 720p, 121 at 1080p, and 283.5 or 483.5 credits for the full twenty. Leaving the duration to the model reserves the twenty-second ceiling up front and settles down to what the clip actually cost, so name a length if you are watching the balance.

Images go in as frames, not as references. One opens the clip, shown pixel for pixel; add a second and the clip travels from the first to the last. There is no numbered reference tray and no subject that can be carried across takes: for that, and for reference video or audio, use Seedance 2.5.

Describe what you want in your own words and Kuro writes the prompt for you. If you would rather write it yourself, this is what FLUX 3 responds to.

FLUX 3 makes the picture and its sound from one prompt, either from nothing or from frames you hand it.

Write plainly, in full sentences. The prompt is interpreted and expanded before generation, so describing the shot the way you would to a person beats terse keyword syntax — this is the one clip model in Kuro that does not reward clipped motion-verb prompts.

Its image channel is a Start/End pair: the first frame is reproduced pixel for pixel as frame one, the last frame ends the clip, and the model fills the middle. With both, describe only what happens between them; repeating the set or the characters in words hands the model a second, differently worded opinion to reconcile.

Without frames, describe what is visible and the action in order, and give the last moment an event.

Sound is generated unless you switch it off. Give it its own sentence — ambience, effects — and put any spoken line in quotes, word for word.

What it can't do

• No reference images, video or audio: the frames are the whole image channel. To hold a face, put it in a frame.
• A last frame needs a first frame.
• 720p or 1080p.
• Two frames too far apart — a different set, a different camera axis — make the take tear or jump near the end. Keep both on the same set and axis; for a large change, make a middle frame and run two takes.
• A gradual change tends to arrive in one hit, because the frames define only a before and an after. Supply an intermediate frame rather than an adverb.
• With an image attached, the provider applies its stricter safety setting.

Use something else when…

• A face or product must be held from reference photos: Kling v3 Omni.
• A reference video or a voice track should drive the take: Seedance 2.5.