Generate & Review (stage 6)

Why parallel

A 30 s 1080p take can take 5–20 minutes. Generation is also stochastic: the same prompt yields different eye-lines, different hand physics, a different third chair. The course rerolled serially and spent days. We fire four at once so that (a) the turn stays alive and (b) we see the distribution of what the prompt produces — which is the only way to tell "prompt problem" from "bad luck".

State the mode, count, model and expected wait before firing.

Scripts (film-pipeline/scripts/)

The helper below forwards images and parameters, not speech attachments. For voiced scenes, complete film-dialogue-voiceover, then use the configured audio-capable MCP/API and verify its submitted speech inputs. Never send an image-only batch as if it contained the approved voices.

SK=<abs path of film-pipeline>/scripts
$SK/batch.sh --model seedance-2-5 --copies 4 \
  --image reference_image=<url> --image reference_image=<url> --image reference_image=<url> \
  --param duration=30 --param resolution=1080p --param ratio=16:9 \
  --out films/<slug>/takes/block-01 films/<slug>/prompts/block-01.txt   # 4 files → max
$SK/poll.sh films/<slug>/takes/block-01            # loops every 30 s; --once for a single pass

batch.sh writes tasks.jsonl and copies the prompt(s) into the take folder so the folder is self-describing. poll.sh downloads <prompt>-c<k>.mp4|png and writes status.json. Run poll.sh in the background for long video batches and tell the user what's cooking.

Drafts: run the first version of a new prompt on seedance-2-0-fast 1080p to validate blocking, then seedance-2-5 at 30 s.

Submission and status — the rules that were paid for

These came from a production where 24 "submitted" tasks turned out to be 24 CLI errors, and a 30-second block came back as 10 seconds without anyone noticing until assembly.

1. A task exists only when the CLI returned a task ID. cohub generate … --async --json must produce a taskRunId. No ID → not submitted, regardless of how many files were written. Count IDs, not files.
2. Check the selected interface's live schema. For the image-only CLI example, inspect cohub generate --help and cohub models show <model> --json; other MCPs/APIs can differ. For voiced takes, also verify audio upload/reference fields, character bindings, accepted formats, URL access, and output-audio settings. Do not invent fields, drop speech inputs, or assume every Seedance wrapper has the same capabilities.
3. Status is read from cohub tasks get <id> --json, field run.status. Do not infer state from files on disk, from process exit codes, or from a custom poller that parses a guessed schema. Results live at run.result.output[].source.url; download from there.
4. Completed ≠ correct. After download, ffprobe the file: duration, video stream, audio stream. A 30 s request that returns 10.08 s is a failed take even though its status is completed; regenerate it with the duration restated in the prompt ("produce the full 30 seconds; hold the final card for the remaining time").
5. Failed tasks are retried per candidate, once, with the failure named. OutputAudioSensitiveContentDetected → remove sung/lyrical audio language from the prompt or move music to the audio pass. Provider 400 → check parameter names against the schema again.
6. Never re-submit a task that already has an ID. Old tasks finish on their own; deleting or duplicating them destroys the comparison you were running.
7. Report only what was verified. "16 tasks submitted, 14 completed, 2 failed, 12 downloaded, 11 pass ffprobe, 1 short" is a status. "All done" is not.

Duration: parameter, buffer, verify

• The requested length goes in --param duration=N. It is never written into the prompt text; the model treats "(15s)" as content and may render a clock, a caption, or simply ignore it.
• Request 2–3 s more than the edit needs. Model start and end are stochastic — the first half-second is often a settle, the last half-second an early fade. The buffer is trimmed in the cut.
• Verify with ffprobe on download. A completed task with the wrong length is a failed take.

Budget mode

When the user says the budget is one pass (no rerolls, no drafts):

1. Say it back: "Budget mode — one take per block, no style check, no drafts."
2. Write every prompt at full detail: all references bound, spatial lock, literal blocking, beats with durations, positive constraints, the count-and-uniqueness line. Detail is the only variance reducer left.
3. Submit all blocks in one parallel wave so the user is not waiting on a chain.
4. Report per block: task ID → status → ffprobe duration → beat checklist pass/fail. Nothing else.
5. If a block fails, do not re-submit automatically. Name the failed beat, identify the missing or contradictory variable, change only that cause, and ask before spending again.

Pre-flight checklist (run before every submit)

• Every character in the cast list has exactly one reference URL, and no other reference contains that character.
• Every location used has its SPATIAL LOCK pasted at the top of the block.
• Every prop that matters has a sheet URL or a text_fallback sentence used verbatim.
• Every screen is a scene-matched mockup of the original capture; language and logo checked.
• Every requested spoken line has a checked recording, stable speaker/voice assignment, measured timing, and an actual supported audio attachment, or an explicitly agreed alternative route.
• The outgoing request retains the speech inputs and supported audio-output settings; narration is distinguished from on-screen dialogue.
• A representative voiced take has passed review before expanding to the remaining batch.
• Every action beat has wind-up → contact → consequence; every effect follows its cause.
• Constraints state "exactly N characters, each appearing once."
• No duration, no seconds, no timestamps inside the prompt text.
• --param duration = edit length + 2–3 s.
• Reference files are accessible to the provider through its supported upload/asset/URL mechanism for the job's lifetime; private audio is not published without authorization.

Long-form assembly

For films longer than one block:

• Each block is generated with the previous block's last frame as first_frame and the same reference_image set. Extract the last frame immediately after download (ffmpeg -sseof -0.1 -i block-N.mp4 -frames:v 1 block-N-last.png), publish it, and only then write block N+1's prompt.
• Splice across hard cuts of different setups; regenerate-merged inside a continuous setup.
• Assemble with ffmpeg -f concat only after every input passes ffprobe. Verify the assembled duration against the sum of the parts.
• Speech is prepared before generation; preserve accepted voices and timing across cuts. Music, effects, and final mixing follow picture lock. Add prepared off-screen narration in the edit when that was the selected route; do not double speech already present in generated clips.

Review — judge by beats, not by "good/bad"

For each take: extract frames (ffmpeg -i c1.mp4 -vf fps=1/3 frames/c1-%02d.png) and read them; for dialogue also extract audio timing if needed (ffmpeg -i c1.mp4 -vn c1.wav). Write takes/block-NN/review.md:

# block-01 — review — <date>
Prompt: block-01.txt (standard ×4, seedance-2-5 1080p 30 s)

| take | 1A wide | 1B Vega CU | 1C Santiago CU | verdict |
|---|---|---|---|---|
| c1 | ✅ seats right | ❌ third chair | ✅ emotion, foot-tap missing | KEEP 1C |
| c2 | ❌ man-spread | ❌ | ❌ | reject |
| c3 | ✅ best geometry | ⚠️ tilt weak | ⚠️ | BASE 1A |
| c4 | ⚠️ | ✅ best Dutch, line clean | ❌ cuts at 13 s | KEEP 1B |

Checklist tells seen: extra chair (2/4) · foot tap missing (4/4) · early cut (1/4)
Pattern → prompt fix: lock furniture count; move foot tap into beat list with rate.
Decision: v2 standard batch for 1A+1C; keep c4 1B as locked.

The checklist (run every take through it)

For every voiced take, listen for exact words, correct speaker and voice, intact endings, and audible level; inspect visible mouths during playback. An audio stream containing only effects fails the dialogue check. Report listening or playback limitations; metadata and transcription alone do not prove lip sync.

identity drift · wardrobe/kit recolour · seat/position vs continuity · counts (people/chairs/balls) · location drift · people vanishing · physics (weightless prop, snapping limb, floaty ball) · eye-tracking stability · emotional entry state vs previous block · missing micro-action · lips moving off-line / speech before a whip lands · shot size change inside a continuous move · text legibility (scribbles?) · early cut · hero's face on a one-off · flat angle (technically fine, no reason to look twice) · screen UI invented

Verdict rules

• 4/4 wrong the same way → the prompt is missing a variable. Rewrite (see film-shot-prompt failure map). Do not reroll.
• 1/4 wrong → fluke. Use the others.
• Every take has one good shot → normal. Decide how to combine (below).
• "Basic but correct" → stop tuning emotion, tune camera.
• Keep backups noted ("c2 ball placement most physical — backup").

LOCK — make the decision explicit

Locking is the thing the user will most want to override later, so it must be legible. Write takes/block-NN/locks.json:

{
  "block": "01",
  "locked": [
    {"shot": "1A", "take": "c3", "span": [0, 6],   "why": "geometry matches script; both seated; tilt legible"},
    {"shot": "1B", "take": "c4", "span": [6, 10],  "why": "best Dutch, clean line, no looking room"},
    {"shot": "1C", "take": "c1", "span": [10, 15], "why": "emotion right; foot tap missing but acceptable"}
  ],
  "rejected": {"c2": "man-spread, flat"},
  "backup": {"1C": "c3"},
  "assembly": "splice | regenerate-merged | single-take",
  "open_debt": ["foot tap not visible in 1C"]
}

Show the user this table in the reply (take → shot → why) and ask them to confirm or swap before assembling. If they change their mind later, the file is the thing they edit.

Combining takes: splice vs regenerate-merged

Two options when the good parts live in different takes:

1. Splice (cut spans from different takes). Cheap, instant. Works when the cut point is a hard cut in the prompt anyway and the shots on either side are different setups (a wide then a CU). It reads unnatural when the two spans are the same setup from different generations — light, grain, head position all shift.
2. Regenerate-merged. Write a v-next prompt that describes the winning qualities of each take explicitly ("1B: her tilt and eye-line as in take c4; 1C: his welling eyes as in take c1, plus the foot tap"), optionally pass the best frames as reference_image, and batch 4 again. Costs one more batch; yields a single coherent take.

Default: splice across hard cuts of different setups; regenerate-merged inside a continuous setup. Write the choice in locks.json.assembly. This is the pipeline author's own finding — spliced same-setup takes look wrong; regenerate instead.

Reuse before regenerate

• Pause beat: the 1–2 s processing pause after a flashback can be a clean span from an already-locked take of the same setup (the course reused the VO take's opening). Look through locks.json of earlier blocks first.
• VO: if only off-screen narration is wrong, use the existing cast voice through film-dialogue-voiceover and retain the picture. Visible dialogue also needs a reviewed mouth-timing repair or a regenerated shot.
• Match cut: reuse the previous block's last frame as first_frame.

Assemble

cd films/<slug>/takes/block-01
ffmpeg -y -i c3.mp4 -ss 0  -t 6 -c:v libx264 -crf 16 -c:a aac s1.mp4
ffmpeg -y -i c4.mp4 -ss 6  -t 4 -c:v libx264 -crf 16 -c:a aac s2.mp4
ffmpeg -y -i c1.mp4 -ss 10 -t 5 -c:v libx264 -crf 16 -c:a aac s3.mp4
printf "file 's1.mp4'\nfile 's2.mp4'\nfile 's3.mp4'\n" > list.txt
ffmpeg -y -f concat -safe 0 -i list.txt -c copy ../../scenes/block-01.mp4
ffmpeg -y -sseof -0.1 -i ../../scenes/block-01.mp4 -frames:v 1 ../../scenes/block-01-last.png
# For narration, mix the prepared speech at its planned offset, preserve ambience,
# and retain the full picture duration; do not trim the film to the voice file.

Publish the block (public-files) and show the link.

Close the block — continuity exit state

Append to continuity.md the exit state of every character as actually seen in the locked take (not as scripted): seat, wardrobe, props in hand, emotional state, eye-line, last-frame path. The next block's prompt opens on this. Then say in one line what was locked and what the next block is.

Audio pass (after picture lock)

• Dialogue repair uses the retained cast voice and exact script through film-dialogue-voiceover; do not recast or clone a generated voice automatically. Keep timing, check visible lip sync, and remove the old speech before replacement. Preserve ambience/effects with stems or a deliberate remix rather than blindly replacing the entire soundtrack.
• Music: suno_music_chirp_fenix, prompt describes genre, tempo, instrumentation and the arc ("quiet piano → swell at 1:10 → resolve"); generate after the cut so the length is known; duck under dialogue in ffmpeg (-filter_complex "[1:a]volume=0.25[m];[0:a][m]amix").

Cost sanity

• Sheets/plates are cheap: always 4.
• Video standard ×4 at 1080p/30 s is the default; max ×16 is confirmed once per film, then not asked again.
• Style check first: block 1 alone in standard; user confirms live-action look and motion before the whole-film wave.
• Lock by index: present takes as block-02 c3 etc. so the user can answer with numbers ("02:c3, 03:c1 0–12 s + c4 12–30 s").
• Drafts on seedance-2-0-fast when a prompt is brand new.