gbax316/make-ai-short

make-ai-short — generative shorts, end to end

The pixels come from a video model; Remotion only composites clips + captions. Proven on blue-man #1 "The Door". Sibling of /make-short (100% TSX) — same discipline, plus three gates TSX never needed: cost stated before generating, character locked from character.json, loop pinned via end-frame conditioning.

Run everything from the repo root. Needs FAL_KEY (clips + images) and ELEVENLABS_API_KEY (voice) in .env.

The three iron rules

1. NEVER generate a character from text twice. ai-shorts/<char>/character.json + character.png is the locked reference. Every still that needs the character is gen_image.py --ref character.png; every clip is image-to-video from a frame that already contains him. If a character sheet doesn't exist yet, making ONE (and getting the user's eye on it) is its own step before anything else.
2. STATE THE COST BEFORE SPENDING IT. Derived, not quoted: Seedance 1080p audio-off = h×w×24×dur/1024 tokens × $1.2/1M ≈ $0.0583/s (fal pricing API validated 2026-07-14 — see ai-shorts/IDEAS.md for the full table + the Wan/Veo-Lite traps). Always generate_audio:false — voice is ElevenLabs, sound is our SFX pipeline.
3. Loop by CONSTRAINT, not luck. Shot 1's true frame 0 (extracted from the clip, not the prompt image) is the loop target. The final shot passes it as end_image_url (Seedance/Kling), and the composition settles onto that exact still over the last ~10 frames (LoopSettle pattern in remotion/src/shots/ai-1/Ai1Door.tsx).

Artifact contract

ai-shorts/<name>/
  script.md · beats.json      — same contract as /make-short (vo[] with real word times)
  character.json              — THE LOCKED REFERENCE (+ video_model block = the picked model)
  shots/NN-slug.png           — per-shot START frame (gen_image.py --ref character.png [--ref scene])
  shots/NN-slug.json          — gen_clip sidecar (model, payload, request id)
  shots/upload-urls.json      — fal storage urls (uploads are reusable across re-rolls)
  voice/ · sfx-plan.json · output/                                          [voice/output gitignored]
remotion/src/shots/ai-N/      — composition + vo.gen.ts
media/projects/<name>/        — THE clip copies for staticFile() (committed — pixels are
                                not reproducible; shots/*.mp4 working copies are gitignored)

Order of operations (voice before pixels)

1. Script — script.md + beats.json. Hook composed at frame 0; thesis should WANT the loop (blue-man #1: "arrival is a myth" → ending = walking toward the same door). ~80-100 words, 2.5-2.7 wps windows. No engagement-CTA outro, ever — end on the payoff, loop is the ending.
2. Voice FIRST — python tools/gen_voice.py --beats ... --emit-ts remotion/src/shots/ai-N/vo.gen.ts (voice and model from .env, or a beats.json "voice" block when a series wants a narrator such as George JBFqnCBsd6RMkjVDRZzb with eleven_v3 tags; act on the LINE CHECK flags as in make-short Stage 4 before any clip is generated). Real word times decide the shot windows; Seedance takes integer 4-12s durations, so round shot lengths UP from windows. Dialogue (two or more characters): beats.json "cast" maps each speaker to its character.json (whose "voice" block is used), each vo line names its speaker in "who", and --pack 0.3 lays the lines end to end from their real lengths instead of squeezing them into estimated windows. One voice per character lives in its character.json, nowhere else. Mouths in dialogue: the video model does not know who is speaking; it makes listeners talk, freezes speakers and stretches a laugh across the next sentence. After the clips, run python tools/mouth_sync.py ai-shorts/<name> (beats.json "shots" gives each clip its global start and length; each character.json "mouth" block gives side, radius, zone, style and keep_above). It erases the model's mouths and draws each character's own from the voice track's loudness. QA it the same way: both faces every half second against who is speaking. A change of action inside one line (laugh, then talk) is two clips, the second started from the first's frame at the cut. Then scan each mouthed clip frame to frame: a spike over four times the clip's median is a cut, a fast gesture or a two-frame morph with ghost features; list morph frames in that shot's "hold" and re-run, so the jump plays as a clean cut.
3. Start frames — one per shot, python tools/gen_image.py --model pro --aspect 9:16 --ref character.png (+ the shot-1 scene still as a 2nd ref when a location/prop must match across shots). QA every PNG before any video call — this is where drift is cheap to fix.
   • Watch hands, clothing details, proportions — name them explicitly in the prompt on re-rolls.
   • Geometry must be walkable (the shot-6 lesson, cost 2 re-rolls): with an end-frame pin, the start frame must let the motion REACH the end state by walking forward — if the end frame's character is CLOSER to camera than the start's, the model spawns a clone instead. Start him larger/nearer than the end state so forward motion shrinks him into it.
4. Clips — state the total derived cost, then python tools/gen_clip.py per shot in parallel (--set image_url=<fal storage url> --set 'duration="N"' --set resolution=1080p --set generate_audio=false, final shot --set end_image_url=<frame0 url>). Upload refs once via bakeoff_clip.upload_ref; keep urls in shots/upload-urls.json. The fal form above is this checkout's default (CLIP_PROVIDER=fal, FAL_VIDEO_MODEL in .env). On kie.ai instead (--provider kie): field names follow the kie model, not fal. For bytedance/seedance-2: --aspect none --set first_frame_url=@ai-shorts/<name>/shots/NN-slug.png --set duration=N --set resolution=1080p --set generate_audio=false, final shot --set last_frame_url=@<frame0.png>. @path uploads the local file (kie keeps it 24h, so upload-urls.json reuse only holds within a day). Cost: kie.ai/pricing, not the fal table; the run prints credits used. bakeoff_clip.py is still fal-only. Prompt every clip with: exact character phrase + "EXACTLY ONE character in the scene for the entire shot" (clones are the #1 failure) + camera move + "no morphing" + style line.
5. Clip QA, frame by frame — start/mid/end strip per clip + a side-by-side of the final clip's last frame vs the loop target. Re-roll failures (state the re-roll cost); a 5s 1080p re-roll is ~$0.29 — cheap next to shipping a clone.
6. Composition — Sequences of OffthreadVideo over ONE timeline, ~8-frame crossfade underlaps (clips are ~1s longer than their windows — that's the slack), Captions + ProgressBar at root, LoopSettle dissolve at the tail. Copy final clips to media/projects/<name>/. npm run gen, frames.mjs at boundaries + heroes, READ every PNG, then render-all --scale=1 and mux voice.wav.
7. SFX — /suggest-sfx flow. Gappy-VO calibration (blue-man #1): cues in VO gaps get no masking and no duck — they need ~7-8 dB LESS than the dense-narration table (gap whooshes -9..-13, not -3..-5). Verify with volumedetect RMS against a speech-mean reference; hero layer may kiss voice level once. Audition mix awaits the user's ear (hard gate).
8. Save back — catalog used_in, beats.json real timings (gen_voice writes them), sidecars committed, cost ledger in the final report.

Done =

character consistent in every frame of every clip (frame-by-frame QA'd) · loop verified (last frame vs frame 0 side-by-side) · captions on real word times · costs stated before each spend and totalled after · SFX audition rendered, awaiting the user's ear · no CTA outro. · tools/check_text.py clean on the spoken and on-screen text · phone copy made with tools/export_phone.py · a dated entry in LESSONS.md (and the skill updated, in general terms, if the lesson changes how it works).