A short tutorial scene can be studied through original casting, several generated takes and a measured timeline edit. ChatCut’s C3 Santiago run exported 14.625 seconds and passed 6/7 MUST checks on October 27, 2026. The framing requirement still failed and preferred audio remained uncertain. This report explains the partial result, rather than claiming an exact recreation of the tutorial or a completed film.
Verified facts and their scope
Last verified: October 28, 2026. The named C3 historical local report supplies the measurements below; public tutorial and provider sources serve different purposes.
| Layer | Finding | Evidence and boundary |
|---|---|---|
| Source tutorial | The study targets the Santiago office scene beginning around 0:53. | Higgsfield tutorial scene, reviewed October 27, 2026; external QA target only. |
| Native model | Seedance 2.0 supports reference-guided generation within its input rules. | BytePlus API, accessed October 27, 2026; not proof of precise framing transfer. |
| ChatCut route in C3 | GPT Image 2.5 Sunburst generated original stills; Seedance 2.0 generated 1080p takes. | C3 Santiago local report, October 27, 2026; no current-production retest. |
| Delivered result | 1920×1080, 24 fps, 351 frames; 6/7 MUSTs passed. | C3 report: framing failed; room-tone/no-music preference unverified. |
Which tutorial and scene are being studied?
The public Santiago tutorial supplies a three-shot reference: a two-person office wide, a psychologist single and a Santiago single. C3 used source frames to judge shot families and pacing. Tutorial footage stayed outside the project and was never a generation reference or a deliverable item.
This substitutes ChatCut’s original still generation, Seedance takes and editable timeline for the source tutorial’s workflow. It does not establish compatibility with the source platform’s presets, project files or model settings. The comparison covers a short scene, not the tutorial’s whole film.
Which assets were generated from scratch?
The run generated 20 character-sheet candidates and 10 office plates, then selected one sheet for each of the two characters and one office plate. The two sheets used one visible face per person to reduce ambiguity. All video references came from these original generated assets.
A repeatable adaptation starts by defining what transfers from the reference: shot order, approximate duration, occupancy, eyelines and camera family. Keep original casting and location decisions explicit. Do not assume that copying descriptive prose will reproduce the source’s exact faces or lens position.
Which shots were harvested from which takes?
Four full takes used the same 15-second prompt. Instead of choosing one entire take, the reviewer compared each shot and assembled the recommended combination.
| Delivered shot | Selected material | Timeline duration at 24 fps |
|---|---|---|
| Two-person wide | Take 3, starting at 0.500 seconds; later rotated −6° and scaled to 118% | 104 frames / 4.333333 seconds |
| Psychologist | Take 2, starting at 4.200 seconds | 100 frames / 4.166667 seconds |
| Santiago | Take 2, starting at 8.900 seconds | 147 frames / 6.125 seconds |
The cut boundaries are frames 104 and 204, or 4.333333 and 8.500 seconds. Adjacent-frame checks found two clean cuts, with no surviving one-frame fragment. Timeline durations add to 351 frames: 14.625 seconds.
These are the actual selected source-in times; they are not claims that generation obeyed exact timestamps. The editor supplied the final pacing.
Where does the export match or miss?
The final scene passed its checks for three ordered shots, two people occupying distinct chairs, the psychologist’s line, bounded character matching, warm appearance and total/per-shot pacing. Manual identity checks covered the contributing takes and original full-take singles; the reviewer had prior project context.
Framing still failed. The wide was lower and more frontal than the source’s high oblique viewpoint. The psychologist shot’s cant pointed in the opposite image direction. Two targeted five-second wide repairs both stayed level and were rejected. An image rotation changed tilt, but could not change the camera’s original position.
The preferred room tone and absence of music were not independently verified. Dialogue transcription and the absence of a separate music asset cannot establish those audio properties. All seven MUST checks would need to pass before calling this acceptance successful.
What did the bounded run cost?
These are settled historical credits from the October 27, 2026 local acceptance cohort, reviewed October 28. They are not a public tariff or a quote for repeating the scene today.
| C3 media | Historical settings and quantity | Settled credits |
|---|---|---|
| Character sheets | GPT Image 2.5 Sunburst, 2K, 20 images | 4.00 |
| Office plates | GPT Image 2.5 Sunburst, 4K, 10 images | 4.00 |
| Main takes | Seedance 2.0, 16:9, 1080p, four 15-second takes, image references, no reference video | 88.92 |
| Rejected wide repairs | Seedance 2.0, 16:9, 1080p, two 5-second takes, image references, no reference video | 14.90 |
| Media total | 30 stills and 6 videos | 111.82 |
The recorded video model was doubao-seedance-2-0-260128; stills used gpt-image-2.5-sunburst. Accepted-run Agent usage added 22.24 credits, for 134.06 with media. Separate readiness usage added 0.76, making 134.82 attributable to this workstream. Values are reported to two decimal places as recorded; the report does not establish a reusable rounding formula. Rejected repairs still cost credits. Obtain a new estimate and compare the settled ledger using the current credit policy.
Apply the method to your own scene
Define the acceptance target
Write the shot families, character positions, eyelines, dialogue ownership and per-shot durations before generating.
Approve original references
Select separate character locks and a location plate. Keep external tutorial material as an attributed review target.
Review multiple takes by shot
Compare identity, camera and motion separately. Record which section of which take is usable.
Assemble and inspect the export
Measure the final cuts and duration, review adjacent frames, and listen to the sound. Carry failed checks into the handoff.
Continue with the existing AI filmmaking guide, model matrix, generation references and workflow failure guide.
Sources
- ChatCut C3 Santiago short-film acceptance report, October 27, 2026, reviewed October 28, 2026: historical local integration run, selected-take review, export measurements, acceptance and settled credit attribution. The internal report and media are not publicly hosted; this article does not claim a current production test.
- Higgsfield Santiago tutorial scene, reviewed October 27, 2026: attributed scene target, not proof of ChatCut output or costs.
- BytePlus generation API, accessed October 27, 2026: native input modes and generation constraints.
