Reference-to-video feeds the model real images or clips (character sheets, location plates, a phone-shot mock) so it carries that exact identity into footage instead of inventing it from text. Seedance 2.0 accepts up to nine images plus a short video and audio clip per generation, and chaining the final frame produces seamless one-take shots without drift.
Reference to video feeds the generation model real images or clips — character sheets, location plates, a phone-shot mock — so it carries that exact identity into the footage instead of inventing one from text. Seedance 2.0 accepts up to nine images plus a short video and audio clip per generation, and chaining each segment's final frames back in produces seamless one-take shots without identity drift.
What a reference actually controls
A reference to video input controls identity: who the character is, what the location looks like, and how the camera should move — things a text prompt can only describe, never guarantee. Reference-to-video provides more context than start-frame/end-frame methods, which hand the model two stills and leave everything between them to chance. It also outperforms extend for one-take workflows, because extend cannot accept character or location references — reference-to-video takes both simultaneously, so the model holds the full identity of the shot, not just its last frame.
Character sheets are your strongest identity reference
A multi-angle character sheet — front, side, profile, back — is the single most reliable identity reference you can attach. One documented 70-second short film kept two characters visually consistent across every scene using character sheets and agent context alone, no LoRA fine-tuning required. Two rules sharpen the sheet itself: include close-up panels, not just wide shots, so small details like scars and accessories survive across generations; and remove objects from the character's hands before generating turnarounds, since held props drift across angles. Locking one character took about $9.78 in a documented production — cheap insurance against re-rolling every downstream shot.
Reference-to-video chaining: continuous takes that don't drift
For continuous takes, chain segments instead of generating one long clip. Clip the final frames of each generated segment and re-upload them to the invideo agent, which attaches that clip to Seedance 2.0 reference-to-video alongside your character sheets and location references to generate the next segment. Because the model receives the full prior clip plus identity references — not just a frame — camera movement, framing, and atmosphere carry across the segment boundary. Two supporting moves: generate a fresh character sheet for each costume-change or prop beat inside the take, and let the invideo agent scout real-world landmark images from the internet to serve as location plates.
Mock-shot and sketch references for blocking the model can't guess
When prompting can't land a POV move or a complex physical arrangement, make the reference yourself. Act the shot out on your phone and upload the footage as a reference video — a documented production cracked a POV shot this way after multiple prompting attempts failed. For multi-character contact setups the model can't visualize from text, hand-sketch the configuration and upload the drawing; the invideo agent feeds it into the image model to produce an accurate character sheet, which then becomes your video reference.
Tag what to adopt and what to ignore
References work better with explicit instructions about what to take and what to leave. Batch reference images by theme — spatial logic in one batch, color theory in another — and tell the invideo agent exactly which quality to adopt from each and which to ignore. For illustrated or animated references, don't feed them in raw: have the invideo agent read their color palette and textures and translate those into the prompt instead. In one production, that extraction method returned hyper-realistic generations at the exact color temperature the references carried.
The agent is the routing layer
You don't manage reference attachment shot by shot — the invideo agent does. invideo is an agentic video creation tool with every current model available, and the invideo agent decides which references to attach per generation and which model to route them to: Seedance 2.0 for reference-to-video continuity, Kling for native multi-shot sequences, Veo where its motion suits the shot. Prompt it the way a director briefs a crew — intent, not parameters — and it couples your character sheets, location plates, and prior clips into each generation.
FAQ
How many references can you attach per generation?
Seedance 2.0 accepts up to nine images plus a short video clip and an audio clip per generation. In practice, a character sheet, one or two location plates, and the previous segment's end clip cover most shots.
Does reference-to-video replace character sheets?
No — it consumes them. The character sheet is the identity source; reference-to-video is the mechanism that carries it into motion. Create a fresh sheet for each appearance change within a continuous sequence.
What's the difference between reference-to-video and extend?
Extend continues a clip from its last frame only. Reference-to-video ingests the prior clip plus character and location references, so identity and atmosphere persist across segments instead of degrading with each continuation.
Sources
Production figures in this article come from documented invideo productions, quoted as recorded.
Watch these to see the techniques in action: