A short checklist for assembling generated shots into a final cut . General montage theory is not covered.
Principles
- Edit and generate in parallel: assemble while shots come out; if an insert is missing, generate it immediately—don't wait for "everything to finish generating."
- Generated footage is by default on the slow side → cut harder than instinct suggests.
- Score and final mix only enter in post; keep generation prompts SFX only.
Cut
| Technique | Method |
|---|---|
| Head and tail | For each shot, first cut about 0.3–0.5s from the head and tail(edges drift and deform most easily) |
| Pacing | Better short than sagging; align dialogue seams using the tail of the previous line / head of the next line (consistent with the prompt seam strategy) |
| Reaction first | After an event, cut first to the face of the person who sees it , not to a repeat of the action itself |
| If it doesn't hold up, replace it | The same prompt log can be rerun; change one line and regenerate; don't force a bad shot to blend on the timeline |
| Empty shots/hands | Actively ask for cutaways; cheaper than forcing cuts on bad performance |
Cleanup (before color grading)
Fix first:face close-ups, hands, fake text/garbled text, extra fingers, boiling textures。
Fix small defects frame by frame; if a whole shot is a write-off → use the final-version prompt and change only one line to regenerate.
Cleanup first, then unify color grading. Each shot carries its own "generation grade"—align color to neighboring shots first, then apply the film-wide look (the look is best already set in direction during the location asset stage).
Sound
- Prefer the lip-sync audio from generation: denoise, unify timbre, place it in the space.
- Only re-record if unusable.
- One continuous ambient bed laid under the whole scene, masking slight image drift between shots.
- Music is added later, following the cut points, and does not enter generation.
Final Cut Acceptance
- Identity/clothing/injuries/prop state continuous across neighboring shots
- No obvious slop (hands, text, flicker)
- No residual "generated-feel BGM"
- Pacing doesn't sag; slow shots have been cut
- Color grading unified within the same scene