Reference to Video
Reference-guided video AI
Attach up to 9 reference images to lock character appearance, product details, or art direction, then describe the shot. Get consistent AI video output guided by your real assets.
Use up to 9 reference images to guide a character, product, or look; generated details may still vary.
Model
Reference images 0/9
Reference videos 0/3
MP4 or MOV, 2–15s each, 15s total, up to 200MB. Requires R2 with a public domain — the model fetches the file over the network.
0/2000
Outputs
Audio
With audio
Synchronized sound, voice and music
Your generated video appears here. Describe a shot and hit Generate.
Reference to video adds source material to a generation request. Attach up to nine images and three video clips, describe the shot, and let the provider use those assets as visual guidance. The result is newly generated, so identity, lighting, motion, and fine detail may still change.
Use references when a prompt alone does not describe the subject or look well enough. A character sheet, product image, or palette board can give the provider more context, but it is not a post-production lock or a guarantee of consistency. Keep the set focused, describe the shot separately, and review the generated take before using it in a sequence.
A reference is an instruction to hold something constant. These are the three that most often decide whether footage is usable.
Supply two or three clear images of the same face — different angles, consistent lighting — and the model keeps that identity across the clip and across separate generations. This is what makes a series of shots feel like one production.
Colourway, proportions, logo placement, materials. Text prompts invent a plausible object; a reference image pins the real one, which is non-negotiable for anything commercial.
Palette, grain, contrast, era, lens character. Feeding in stills that share a look transfers that look, so a batch of clips reads as a coherent campaign instead of a mood-board accident.
The instinct is to upload everything you have. Resist it — references compete with each other, and nine conflicting images produce a worse result than three agreeing ones. Pick images that share lighting logic and are unambiguous about the thing you want held: for a face, two or three angles of the same person under similar light; for a product, a clean pack shot plus one detail; for a look, stills that genuinely share a grade rather than merely a vibe. Crop out anything you do not want carried, because the model has no way of knowing that the busy background was incidental.
Video references work differently from image references, and it is worth being deliberate about which you reach for. An image says *what things look like*; a video clip says *how things move* — pacing, camera behaviour, the rhythm of an action. Attaching a clip whose motion you admire lets you apply that movement to a scene the clip never contained. Combined budgets apply: up to three reference videos totalling no more than fifteen seconds, so choose the segment that carries the move rather than uploading a whole take.
Conflicting lighting or contradictory angles force the model to average. Consistency between your references is what produces consistency in the output.
Use stills to fix identity and art direction; use clips to borrow a camera move or an action’s rhythm. Mixing both in one request is where this tool is strongest.
Model specifications
Published limits, not marketing rounding. Every number below is enforced by the generator before your credits are reserved, so a job that would exceed a limit is rejected instead of failing halfway.
The model is ByteDance’s. What FrameAI adds is the part that decides whether it is usable at work: honest limits, honest billing, and output you are allowed to ship.
Text, up to nine reference images and up to three reference video clips are fused in one generation, not stitched afterwards. There is no separate text-to-video mode to switch into — you add whatever material you have and the model reconciles it.
Seedance 2.0 writes footsteps, room tone, impacts and score alongside the frames, locked to the action. Most models hand you a silent clip and leave the sound design to you; here the export is already finished.
One prompt covers a 9:16 vertical cut for Reels, a 16:9 master for YouTube and a 21:9 anamorphic frame for a title sequence. Resolution runs from 480p up to 4K, so the same take can serve a cinema-shaped frame or a phone.
The exact credit cost is calculated from your duration, resolution and ratio and shown before you press generate — never after. If a generation fails on the provider side, the reserved credits are returned automatically.
Seedance 2.0 Fast and Seedance 2.0 Mini exist for the twenty drafts nobody sees. Rough a shot at 480p for a fraction of the credits, then re-run the take that worked at full resolution on the flagship model.
Paid-plan output may be used commercially under the Terms of Service. Free-credit output is for evaluation only.
Set up the references once and every subsequent shot in the series inherits them.
Upload up to nine images and three clips. Crop each one to the thing you want held, and drop anything that fights the others on lighting or angle.
The references already say what things look like. Your prompt should say what happens and how the camera behaves — re-describing the reference wastes the instruction.
Keep the same reference set and change only the prompt to produce a second, third and fourth shot that all belong to the same world.
Each of these fails immediately on a text-only model, for the same reason: something has to stay exactly itself.
A spokesperson or mascot appearing across a dozen clips has to be the same person every time. A locked character reference makes a campaign out of what would otherwise be a dozen strangers.
Hundreds of SKUs, each needing motion, each needing to be accurate. Swap the product reference, keep the prompt and the art direction, and the whole catalogue matches.
Recurring characters and locations across episodes need continuity that a prompt cannot guarantee. References give a series its through-line.
It is video generation that accepts source images or videos as visual guidance. You can attach up to nine reference images and three reference video clips, then describe the shot you want to generate.
Up to nine images and up to three video clips in a single request, with the combined reference video length capped at fifteen seconds. More is not automatically better — references that disagree with each other degrade the result.
There is no fixed consistency guarantee. Two or three clear references of the same subject under compatible lighting can provide better guidance, while extreme angles, occlusion, or conflicting references can increase drift.
Image references fix appearance — identity, colour, materials, art direction. Video references contribute motion — camera behaviour, pacing, the rhythm of an action. You can use both in the same request, and doing so is where the tool is strongest.
Yes. Keeping the reference set fixed and changing only the shot prompt is a sensible way to guide a related set of clips, but inspect each result because separate generations can still differ.
Only with that person’s consent. The model has no way to verify permission, so the responsibility sits with you — and generating identifiable people without consent is not covered by your commercial licence.
Each tool is the same Seedance 2.0 model pointed at a different job. Move between them freely — your credits, history and exports are shared.
Write one line, pick a ratio, press generate. No editing suite, no render farm, no post-production pass — a finished clip with sound, ready to download.
