AI Video · Reference to Video
Keep characters consistent across video shots
Hold the same face, wardrobe, and room while the camera and the action change from shot to shot.
What reference-to-video is holding still
Reference-to-video is for a sequence, not a one-off. You supply stills or clips of the person, the costume, and the place, then ask for a new camera or a new action. Ad sequences, series shorts, and brand stories fall apart when shot two recasts the lead. References are how you refuse that recast. The prompt may change the blocking. It should not change who showed up.
Camera changes are allowed. A wider shot, a profile, a hand entering frame—those are coverage. A new jaw, a new jacket, or a room that swapped paint color is not coverage. Light and wardrobe count as continuity, not as style seasoning. If you only stare at the face and let the kitchen become a warehouse, the series still looks like it moved to another set.
Image to video animates one still. It does not promise the next setup. Text to video will invent a plausible stranger. Video extend continues one clip rather than cutting to a new angle on purpose. Reference-to-video is the path between those: a cut that is new, and a cast that is not.
Prove it before you storyboard ten shots. Generate two clips from the same references and compare the face, the clothes, and the room. If either drifts, fix the references or narrow the action. Do not “solve” drift by writing a longer paragraph that says consistent. The pictures are the evidence.
Use this after you know the cast. Discovering a character belongs in text to image or text to video. Repeating them belongs here. A brand film that changes actors between the hook and the product shot is not a style. It is a continuity error.
What references keep in motion
The same face across cuts
Age, haircut, and features should survive a new angle. That is the reason to use references.
Wardrobe and palette
The coat and the accent color are continuity. Do not restyle them mid-sequence.
A room that stays a room
Light and set dressing may be seen from a new side. They should not be replaced.
Action and camera may change
A new verb or a new lens is the shot. The cast is not the shot.
A two-clip gate
If two clips do not match, do not generate the other eight.
Series, ads, and brand stories
Anywhere a viewer would notice a recast. One viral clip does not need this. A sequence does.
How to shoot a matching next shot
- 1
Gather the cast references
Clear face, costume, and a view of the room. Skip images that contradict each other.
- 2
Write the new shot only
Action, camera, duration. Then repeat who must stay: same person, same clothes, same place.
- 3
Generate two versions
Compare them to each other and to the references before you fall in love with a move.
- 4
Reject drift early
A prettier clip with a new face is a failed reference-to-video take.
- 5
Change coverage, not identity
The next prompt can move the camera. It cannot introduce a new jacket “for variety. ”
- 6
Extend only after the cut works
Length belongs to video extend. Do not ask a continuity pass to also invent extra seconds of a new idea.
Copy-ready reference-to-video prompts
Same courier, wider shot
Same person, gray jacket, and stone stoop as the references. Wider camera, they knock once and wait. Five seconds, overcast. Do not change face, clothes, or building color. No readable text.
Same botanist, new angle
Lock the referenced face, haircut, and olive coat. New shot: profile, they look down at a plant, greenhouse unchanged. Four seconds, soft daylight. No new accessories.
Product in hand, same kitchen
Same person and the same sage mug as the references. Same kitchen light and cabinets. They lift the mug slightly and pause. Static camera, four seconds. Do not recast or move rooms.
Continuity check, almost still
Same cast and room as the references. Medium shot, they breathe and blink only. Three seconds. This is a match test, not a new scene and not a costume change.
Sequences that stay one world
Agree the references first
Two photos of different days will average into a third person. Pick one look.
Say the room, not just the hero
Reference-to-video drifts in the background when the prompt only mentions the face.
Do not add variety in the costume
Variety belongs in the action. A new jacket is a new film.
Gate on shot two
If shot two matches, continue. If it does not, more shots will only multiply the error.
Who needs matching shots
Ad teams cutting a sequence
The spokesperson has to be the same human from hook to offer.
Series producers
Episode shots cannot recast the lead because the prompt got bored.
Brand storytellers
The place and the clothes are recognition, same as the logo.
Editors who hated a one-off
Image to video gave them one moving still. The next angle needs memory.
Directors testing an engine
They judge Seedance or another model on continuity, not on a single spectacular clip.
Reference to video FAQ
How is this different from image to video?
Image to video animates one still. Reference-to-video keeps a character and setting consistent when you need another shot.
Can the camera change?
Yes. Coverage can change. The person, wardrobe, and room should not.
What if only the face matches and the room does not?
That is still a break. Light and set are part of continuity. Lock them in the references and in the sentence.
Why did I get a similar stranger?
The references were unclear or contradictory, or the prompt asked for a new age, outfit, or location. Narrow the change.
Should every short use this?
No. A single moving poster can stay on image to video. Use references when there is a cut to protect.
Can I discover the character here?
You will fight the tool. Find the face first, then lock it. Reference-to-video repeats a decision. It should not invent one.
Does a video model guarantee the match?
No. Two-clip comparison is part of the job, whichever engine you pick.
Related guides
aboutPage.cta.text
aboutPage.cta.button