| date | learning | source |
|---|---|---|
| 2026-10-06 | Registry role 'Motion transfer / recast' has no entry point. kling_motion in scene_parser.py/test_pipeline_fixes.py is only an i2v prompt-text hint. | knowledge/ad-formats/_MODEL-REGISTRY.md; grep scripts 2026-10-06 |
| 2026-10-06 | fal minimax/h3-max/recast keeps the source's motion, camera, cuts AND audio (video_url 5-30 s, no shot >15 s; reference_image_urls one per person; 768P/1080P; $0.30/s 768P, $0.45/s 1080P per page 10-06). Because the set and audio survive, recast is for own or licensed footage only. | fal OpenAPI + price page 2026-10-06; output/model-catalog/fal-2026-10-06.json |
| 2026-10-06 | fal Kling v3 motion-control renders characters AND background from image_url; video_url only drives motion (orientation image <=10 s, video <=30 s); keep_original_sound defaults TRUE, so always set it false for third-party sources. | fal OpenAPI schema fal-ai/kling-video/v3/standard/motion-control 2026-10-06 |
| 2026-10-06 | Charswap precedent: 24 of 25 live ads edited ONE source video, 4 of 28 got delivery. The competitor-rip source leaked a tracker-module prior into our band (5 of 7 visual-QA fails). Spread sources and never let competitor footage into a model. | scripts/charswap_source_refs.py, scripts/charswap_overnight.py |
| 2026-10-06 | Higgsfield CLI exposes hf_mult_motion_control (>=1 image + exactly 1 video ref), kling_video_edit, seedance_2_0 video refs, and seedance_2_5 omni_reference/video_edit. All paid since 9-09, untested for motion transfer. Run 'higgsfield generate cost' (free) first. | higgsfield model get, 2026-10-06 |
| 2026-10-06 | davidaistar motion-swap rights rule: change the character AND the background; swapping only the face keeps the creator's identity and drew complaints. His demos still kept the creator's audio; ours never do. | analysis/gemini_Rf1uHxSWpjU.md, analysis/ugc_seedance.md (Rqim7SlMtZg) |
| 2026-10-06 | Audit: mode R must not recast segments with the band in a hand or on a wrist (recast regenerates hands; generated hand on band = 3-4/10, banned); those segments stay real footage. Source-type names aligned to format-ad-clone (own-winner / organic-viral; licensed added here). Creator footage needs Fish to confirm usage terms cover AI modification. | format-ad-clone SKILL.md §5; project_dawn_yapper_2026-10-04; ugc-engine/sops/04-outreach-sop.md |
| 2026-10-06 | Motion-control prompt names the one thing borrowed (gesture + upper-body timing) and what the target image fixes; unstated, the model also borrows the clip's framing, styling and pace. Source = one uninterrupted take; montages are not motion refs. | starpop how-to-use-kling-3-0-to-make-ads step 10; how-to-use-seedance-2-0-to-make-ads |
| 2026-10-06 | H3 Recast proven 10-06 (Fish: 'recast is fine'): poses, framing and cut timing preserved on a cartoon clip; $1.56 for 5.2 s. | output/model-proofs-2026-10-06/README.md |
| 2026-10-06 | Fish 10-06: creator footage may be AI-edited/recast (rights gate now allows our creators' footage). | Fish chat 10-06 |
| 2026-10-06 | Photoreal routing moved to ugc-omni (Fish 10-06): photoreal target images = ugc-omni G1 avatars from a real creator frame; spoken photoreal lines and mode R refresh of a talking-head UGC winner = a ugc-omni job re-performing the winner transcript (omni.py mine words.json), not recast (recast would keep the old voice on a new face). Band segments: never recast (no product gate); real footage or gated omni b-roll (pick --product). Example in references/omni-job.example.json. | skills/ugc-omni/SKILL.md; Fish 10-06 |
format-motion-transfer
Motion transfer / actor swap: take only the MOTION of a source clip (choreography, camera move, comedic timing) and re-perform it with OUR cast, OUR set and OUR audio, either as a trend-gag hook spliced onto a paying Dawn body or as a new-cast fatigue refresh of our own winner.