AI Previsualization to Video: Editable Blocking and Measured Cuts
AI previsualization can keep actor positions and camera paths editable before you generate finished-looking footage. Approve a guide, bind separate appearance references, review a structural take and edit the chosen result. ChatCut’s historical B diner run exported 30 seconds and passed 15/17 checks. The useful side-view proof failed and speaker/no-music audio remained unverified, so the run demonstrates partial acceptance rather than geometry-perfect transfer.
Verified facts and their scope
Last verified: October 28, 2026. The measurements below come from the B diner local acceptance report dated October 27, 2026, named in Sources. They do not certify current deployed production.
| Layer | Fact | Evidence and boundary |
|---|---|---|
| Native model | Seedance 2.5 accepts reference video and images and supports generation up to 30 seconds. | BytePlus generation API, accessed October 27, 2026; native support is separate from a particular ChatCut route. |
| ChatCut and configured route | B submitted the approved 30-second guide plus six one-face casting sheets and one diner plate to Seedance 2.5, first at 720p and then 1080p, 21:9. | B diner local report, October 27, 2026; this was a structural review followed by a new take, not a native Draft→final claim. |
| Editable source | Table, six color-coded seated actors, floor, wall, light and cameras remained in an editable previs scene. | B report: plan view was readable; the side review was occluded by the enclosing wall. |
| Acceptance | 30-second export; 15/17 checks passed. | B report: side-view proof failed; speaker assignment, room tone and no-music preference were unverified. |
The model matrix and generation reference guide explain model selection and input boundaries.
How is blocking kept editable?
Keep the source scene rather than flattening the plan into a single image. Give each actor a stable color and position, then describe the mapping from proxy color to the finished character’s wardrobe. Retain table placement, chair positions, camera paths and cut times as separate decisions.
In B, six color-coded seated actors surrounded a diner table. The approved guide established who appeared near the camera, who faced it and when a shoulder left the frame. If a seat changes, recheck every shot affected by that shared geometry; fixing one camera can break another.
A useful plan view and a readable side view answer different questions. B’s enclosing wall blocked the side review, so its archived source did not satisfy the useful-side-view requirement. For your own review, create a cutaway or move the review camera so actors and heights are visible, then recheck the clean guide.
How do cameras become a guide?
Render the approved camera sequence as a clean video without grids, labels or controls. In B, the guide ran for 720 frames at 24 fps, with cuts at frames 336, 408 and 468. That is 14, 17 and 19.5 seconds. Approve the blocking before spending on a finished-looking take.
Bind the guide to camera motion, cuts and blocking; bind character sheets to appearance and the location plate to décor. This division matters: a proxy’s color should identify a role, rather than turn into the person’s skin or become an unexplained prop.
The source previs tutorial is an attributed reference for the method. The reported ChatCut measurements come from its own local B run, not from that video.
What visual constraints survived generation?
B reviewed a 720p structural take before producing a 1080p take. The final review found the four planned blocking families: the opening orbit with a shoulder at frame right, then the successive near/facing pairs and the final unblocked face. Six distinct seated men appeared without a seventh, and background diner activity was visible.
These were family matches. Spacing differed, a closing nod was faint, and identity matching was manual without blind negative controls. Several characters shared a similar hair family. The result does not establish exact face, décor, grain or geometry transfer.
What did the 30-second export contain?
| Stage | Actual dimensions and timing | What to read from it |
|---|---|---|
| Approved guide | 2520×1080; 24 fps; 720 frames; 30.000 seconds | Camera and blocking reference, not generated picture. |
| Generated 1080p take | 2206×946; 30.041667 seconds of picture; 30.080-second container | The 1080p setting did not mean an encoded 1080-pixel height at 21:9. |
| Final export | 2520×1080; 24 fps; 30.000 seconds | Delivery dimensions differ from the native generated pixels. Export size does not prove extra generated detail. |
The raw final take and export cut at frames 336, 407 and 467: 14.000, 16.958333 and 19.458333 seconds. Compared with the guide, the differences were 0, −1 and −1 frames. Exactly three cuts were found and adjacent frames were reviewed. Preserve both guide and final timings so a reviewer can see what survived.
Reproduce the review process
Build and inspect the editable scene
Set actors, props and camera paths; inspect readable plan and side views before locking blocking.
Approve a clean guide
Review the complete camera sequence and every cut, then export a guide without editor overlays.
Bind appearance separately
Approve one-face sheets for the characters and a location plate. State what each reference controls.
Review structure before the next take
Check blocking, occupancy and camera families at the lower-cost review stage. Quote any next generation separately.
Assemble and measure delivery
Trim on the timeline, measure dimensions and cut timings, inspect seam frames and listen to exported sound.
Which tests failed?
The run passed 15/17 scene checks. A useful side-view proof failed because the wall hid the actors. The sound requirement also did not pass: the transcript recovered six lines, but speech timing crossed cuts and speaker assignment, room tone and absence of music were not independently verified. A transcription annotation is not proof of music either way.
The separate workflow review also found late Look decisions and a missing production-design step. Keep your approved blocking, reference bindings and current prompt aligned so a later generation does not use an obsolete seating plan. This example does not validate every previs editing or export format.
Check workflow failures when a take misses a constraint, and use the current credit policy for estimates. The 720p and 1080p B takes were separately generated and charged; preview review does not imply a free final take.
Sources
- ChatCut B diner previs-to-video acceptance report, October 27, 2026, reviewed October 28, 2026: historical local integration run, archived editable scene, guide and generated-take review, measured cuts and export. The internal report and media are not publicly hosted. Current production and all export formats were not retested.
- Previs tutorial, reviewed October 27, 2026: attributed method reference; not evidence for the B run’s numeric results.
- BytePlus generation API, accessed October 27, 2026: native Seedance 2.5 duration and reference modes.
