FLUX 3 Video

FLUX 3 Video is Black Forest Labs' video model. It generates clips of 5 to 20 seconds with native audio, working from a text prompt, a reference image, or both. On Voor it runs through the generator on this page — no separate account or setup required. The model handles text-to-video and image-to-video in the same interface, and the audio track is generated alongside the picture rather than added in a separate pass. That single-pass design means environmental sounds, dialogue, and effects land in sync with the visible action by default, which removes the manual alignment step that makes layered audio workflows tedious.

Yes, and audio is on by default. Dialogue, ambience, and effects are produced in the same pass as the image rather than layered on after, so they land in sync with visible action. Write sound into the brief the way you write physical action: name the source, quote any spoken line, and tie each effect to something happening on screen. The model does not invent a musical score unless you ask for one, so if you want a clean sound bed of just ambience and foley, simply leave music out of the prompt. Silence is also a valid choice — stating 'no ambient sound in the final second' gives you a clean audio tail for editing.

Anywhere from 5 to 20 seconds, in one-second steps. Cost scales with duration, so pick the length from the edit you actually need rather than defaulting to the maximum. Twenty seconds only pays off when the clip has somewhere to go — an opening state, one cause-and-effect event, and a defined ending that an editor can cut on. Past about twelve seconds, the brief needs a real arc with progression; a static scene that simply holds for twenty seconds wastes the extra duration and the extra credits. Draft at a shorter duration first to confirm the action reads, then extend once you know the composition works.

720p or 1080p, in 21:9, 2:1, 16:9, 4:3, 1:1, 3:4, or 9:16 — or auto, which lets the model choose. Set the ratio before writing the brief: it decides where motion can happen and where overlaid text will sit, and changing it after a render means starting over because composition is not ratio-independent. 1080p costs noticeably more per second than 720p, so the practical workflow is to iterate at 720p until the action, framing, and sound all work, then promote the winning take to 1080p for the final render. That approach typically costs less than getting a single 1080p generation wrong twice.

Draft renders the same brief at 720p for roughly a third of the credit cost. Use it to test whether an idea holds up — framing, pacing, whether the action reads, whether the sound cues land where they should — then rerun the winning brief at full quality. Iterating on drafts and promoting once is almost always cheaper than getting a full render wrong twice. Draft mode is also useful for exploring unfamiliar subjects: if you have never prompted a specific type of motion before — glass shattering, fabric tearing, liquid pouring — a draft tells you whether the model can handle the physics at all before you commit full credits to the attempt.

Yes. Image-to-video takes a source image plus a prompt describing what should move. It works best when the subject is clear and the background is not competing for the model's attention. Describe the motion you want rather than re-describing the picture — the model can already see the image and will treat your text as instructions about movement, camera, and sound, not a replacement caption. Keep the source image clean and well-composed; artefacts, heavy compression, or ambiguous edges in the input tend to amplify during animation. Crop and retouch the still before uploading rather than hoping the motion pass will smooth things over.

The opening frame, one physical action with a clear cause and visible effect, camera position and any movement, the sound environment and its sources, and what the last frame looks like. Concrete physical verbs beat mood words: say 'rain beads on the jacket and the camera tracks at shoulder height' rather than 'it should feel cinematic'. End with a settling instruction — 'camera holds, subject faces left, one second of room tone' — so the clip has a clean tail for editing. Prompts that read like a shot list produce better results than prompts that read like a poem, because the model can execute specific spatial instructions and has to guess at abstract emotional ones.

Change one variable at a time. Keep the subject, duration, and closing frame fixed while you adjust camera distance, lighting direction, or a single audio element. Rewriting the whole brief between runs makes it impossible to tell what actually helped — and every run costs credits. Record the exact prompt and a specific reject reason for each failed take. The reject log is the most valuable tool in the iteration process: after three or four runs, patterns emerge — the model consistently drops an object, or lighting from a certain direction causes shadow artefacts — and those patterns tell you what to change next. Without the log, you are guessing on each attempt.

Yes — short product reveals, material studies, unboxing sequences, and social hooks all fit the model's strengths. Put product geometry, label placement, brand colour, and logo treatment in the continuity rules so they survive the motion. The generated footage still needs the same review as filmed footage: product accuracy, advertising claims, trademark clearance, likeness rights for any recognisable person, and whatever the destination channel requires for compliance. Keep the prompt, source references, generated output, and approval decision archived together with the published asset so you can trace provenance, regenerate a variant, or answer a rights inquiry months later.