Consistent AI characters need approved appearance references and review of each shot. In ChatCut’s historical C3 Santiago run on October 27, 2026, selected takes matched the two original character locks in bounded manual checks. The 14.625-second scene passed 6/7 MUST checks; framing failed and preferred audio remained uncertain. This is a useful continuity method with a small measured example, not an identity guarantee.
Verified facts and their scope
Last verified: October 28, 2026. Run measurements below come from the named historical report in Sources, reviewed on that date. They describe a local integration build, not a fresh test of deployed production.
| Layer | What the evidence supports | Source and limit |
|---|---|---|
| Native model | Seedance 2.0 accepts appearance references, subject to its input-mode limits. A reference is guidance, not a guarantee of the same person. | BytePlus generation API, accessed October 27, 2026. |
| ChatCut run and configured route | C3 used original character sheets and an office plate, then Seedance 2.0 at 1080p. No tutorial images entered generation. | C3 Santiago local acceptance report, October 27, 2026. Current production availability was not retested. |
| Identity review | Selected wide and singles matched their own locks; shuffled controls recorded 8/8 positive matches and negative rejections across the original full-take singles. | C3 manual review; reviewer had prior context. This was not independent biometric verification. |
| Delivered scene | 351 frames at 24 fps: 14.625 seconds, 1920×1080. Six of seven MUST checks passed; framing failed. | C3 export measurement. Room tone and absence of music remained unverified. |
See the model capability matrix before choosing a model and the reference guide before preparing inputs.
What should a character lock contain?
A useful lock names one character and binds that name to one approved reference. Record facial appearance, hair, wardrobe and any features that must survive changes of pose or lighting. Keep the location reference separate so the model can distinguish a person’s appearance from the room around them.
C3 generated separate candidates for Santiago and the psychologist, then selected a sheet for each person. The selected sheets retained one visible face; faces in other body views were removed. This reduces competing facial cues. It does not prove that every future angle will preserve the same face.
Use an original or authorized reference you can reuse. An unfamiliar shot should be judged against that approved lock, rather than against a face that has already drifted in a previous take.
How are references reused across shots?
Keep the same approved reference bound to the same character throughout the shotlist. Describe the action, eyeline and framing separately for each shot. In C3, every video used the generated locks; the tutorial was an external review target, not an input asset.
A practical reference note can be this simple:
Character A: use the approved Character A sheet for appearance.
Character B: use the approved Character B sheet for appearance.
Office: use the approved office plate for the room.
Shot 2 shows Character B alone, looking toward Character A off-screen.This is a writing pattern, not the exact tested prompt. Check the selected model’s accepted input combination before submitting it. Do not assume that sharing a seed guarantees identity across shots; keep the original reference and review the result.
What happens when a character leaves the frame?
On re-entry, bind the original lock again and inspect the returning face, hair, wardrobe and scale. A model can preserve the room while changing a face, or preserve a face while moving the character to the wrong chair. Check both identity and scene geography.
C3 covered a two-person wide followed by individual shots. It did not establish continuity through a long disappearance, new age, costume change or many later scenes. Treat those as new acceptance cases. If a returning face fails, reject that section or generate a targeted replacement; do not silently use it as the next identity reference.
How was continuity checked?
Review the first, middle and last frame of each relevant shot, plus frames at the cuts. Compare each person to their own sheet and to the other person’s sheet. Negative controls matter: a check that accepts every candidate cannot show that it distinguishes the two identities.
C3 saved shuffled identity verdicts before revealing the answer key. The report records passing positive and negative controls in its sampled scope. The reviewer still knew the project, and unused repair takes were outside the delivered continuity claim. These limits should travel with the result.
What still drifts?
Identity continuity did not fix the camera. C3’s opening remained lower and more frontal than the source’s high oblique view. The psychologist’s pronounced camera cant ran in the opposite direction. Rotating the image could tilt it but could not create the missing viewpoint.
Sound is another separate check. Recognized dialogue does not establish room tone, speaker quality or the absence of music. Build a review sheet with separate rows for identity, occupancy, eyelines, framing, cut timing and sound; passing one row does not pass the rest.
For the broader process, use the existing AI filmmaking workflow. For retries and rejected outputs, see workflow failures; check current credit policy before commissioning more takes.
Frequently asked questions
What happens when a character leaves the frame?
The re-entry question is grounded in a public discussion about off-screen character continuity, reviewed October 28, 2026. That discussion supplies the user question, not proof of ChatCut capabilities.
Sources
- ChatCut C3 Santiago short-film acceptance report, October 27, 2026, reviewed October 28, 2026: historical local integration run; original generated locks, manual shuffled identity controls, selected-take harvest and measured export. The internal report and media are not publicly hosted. Its results do not certify current production or arbitrary characters.
- Seedance generation API, BytePlus, accessed October 27, 2026: native reference modes and seed-control scope.
- Santiago tutorial scene, reviewed in the supplied research October 27, 2026: attributed visual review target only; it is not evidence for ChatCut’s measured results.
