| date | learning | source |
|---|---|---|
| 2026-08-16 | Fish: bergamot 3D 'is not even claymation'. True clay = handmade stop-motion, matte plasticine, fingerprints, miniature diorama, tilt-shift, dot-eyed figures; negatives against smooth 3D/Pixar/glossy; anchor frame on every ref gen. | project_claymation_batch_2026-08-14.md |
| 2026-09-19 | Bluey flat-2D rejected: the image model always drew outlines; disney3d replaced it. Canvases now run GPT Image 2.5, so a re-open needs a one-sheet proof first. | project_song_lane_2026-09-19.md |
| 2026-10-04 | Anchors: disney3d = exhibit A frames (paying parent), skeleton = Pilloway (100 ads / 629 dups), clay = our own 8-14 final (no outside narrative-clay winner after two search rounds). Ad-copy search can't find visual styles; transcript openers + Gemini on 3 frames can. | output/dawn-iterate-2026-10-04/video/style-refs/README.md |
| 2026-10-06 | canvas.py STYLE_BLOCKS['claymation'] is the generic 'cinematic claymation' line and reskin.py writes it into re-skin packs; the Fish 8-16 stop-motion block lives only in the 10-04 sheet prompts. Override style_block on every clay pack. | scripts/cartoon_h3/canvas.py, scripts/cartoon_h3/reskin.py |
| 2026-10-06 | Every batch seam is a hard cut: refs_for_batch carries nothing forward and assemble concatenates. In-batch transitions work via transition_in -> 'Transition:' lines. Last-frame chaining is NOT BUILT. | scripts/cartoon_h3/run.py, h3.py, storyboard.py |
| 2026-10-06 | Zak-D rule kept for Dawn: overhead/observer framing can't name her in lines 1-2 and every Dawn winner is identification; the davidaistar Zach-D video shows workflow only, no results. Take the continuous-camera grammar, not the framing. | skills/cartoon-pipeline/SKILL.md, analysis/gemini_IolBFaQIrl4.md |
| 2026-10-06 | Best image model is style-dependent per davidaistar (Seedream for Pixar/Zach-D, a different model for paper); paper only needs pans/zooms, so it belongs on the cheapest simple-motion route (tier C). Both routes NOT BUILT here. | analysis/diff_animation_styles.md |
| 2026-10-06 | Audit: as-004 is stale. reskin.py PROVEN_BLOCK_PACKS now copies claymation's style_block from the nottardy 10-04 pack; the canonical clay block is that pack line, not the condensed procedure text. Other styles still get canvas.py's generic block. H3 cost: h3.py FAL_PRICE (768P 0.05 < 480P 0.08) contradicts its own $1.20/15 s note; budget $0.08/s. New style = one tier-A draft batch first (core gate 6b). | scripts/cartoon_h3/reskin.py:41-90, h3.py:31-33, crosscutting audit 10-06 |
| 2026-10-06 | Clay holds only in bright warm practical light; dusk/low-key turns it plastic or blue-skinned. Dawn wake shots keep a warm lamp in frame and pin skin warm against any cold source; hard objects written as clay imitations. | starpop ai-claymation-prompts, ai-claymation-ad-examples |
| 2026-10-06 | Emotion words ('worried', 'exhausted') render neutral faces; write face mechanics (brows, mouth, eyelids, cheek on palm). Unnamed crowds/witness rooms render as mush: describe each background face. | starpop ai-claymation-prompts, ai-claymation-ad-examples |
| 2026-10-06 | Zach-D stage = an empty eye-level plate attached to every canvas (not an isometric layout sheet, not described in words); off-stage ≤~5 s; hard cuts are the tell (originals 0-2 per ~50 s, best copies ~1/8 s). Chained last frames are start refs only, never identity refs. | starpop zack-d-films-ad-format-explained, how-to-make-zack-d-films-style-ads |
| 2026-10-06 | Built 10-06 (supersedes as-005): run.py --chain-frames hands N-1's frame at its span to same-scene/env seams as a START ref after the canvas (h3.py label branch), drops the lowest env sheet past 5 images, runs chains in order, cascades --regen, logs STALE. Auto seams (no pack chain: key). Proof pending. --tier-models routes H3 models per tier. | scripts/cartoon_h3/run.py, h3.py, tiers.py (Builder B) |
workflow-animation-styles
Choose, lock and run ONE animation style for any Dawn cartoon format (claymation, paper/cut-paper, disney3d, pixar, skeleton, bluey 2D, bergamot, plus the Zach-D continuous-camera grammar): style anchor from a proven ad, character + isometric floor-plan environment + product sheets in that style, the per-style model ROUTE (image role, video role, shot tier), in-batch transitions and the last-frame chaining spec, and style re-skins of proven ads via reskin.py.