# Shin-chan drinks from the gourd bottle

## Appearance revision, version 2

The first generic procedural character failed the user's likeness review. Do not treat passing animation/browser tests as evidence that the character matches the photographs.

The revised pipeline uses the three original figurine photos to generate one transparent standing reference with OpenAI's imagegen tool (the tool did not expose an exact model version). Its face is projected with UVs onto a manually shaped 3D head with front and back surfaces. Forehead, cheeks and jaw have real curved geometry; a bounded part of the eyebrow region receives relief from image darkness. The body, clothing, ears, hands and shoes use code-built geometry. Continuous curved sleeves replace the long capsule arms and exposed ball elbows. Arm reach and bottle grip height were adjusted together, without stretching the product geometry.

Microsoft's official TRELLIS.2 Space was tested for image-to-3D. No GLB was obtained: preview transport did not complete and the follow-up was blocked by anonymous ZeroGPU quota. TRELLIS.2 is not an asset source for the final scene. Preserve the attempt logs under `qa/rebuild/outputs/`.

Front likeness is improved by the reference-conditioned facial image, but it is not exact photogrammetry: the prepared AI image can itself alter the toy, the side/back and body remain manually approximated, and the bottle has not been measured or scanned. The result remains a controllable character-motion prototype rather than a precise product digital twin. Clearly explain image texture versus geometry on the page.

Final render: the fixed image texture and fixed 3D model are reused for the full 20 seconds. No AI video model or TTS is used. English frame text, no audio. The previous version is retained under `qa/character-v1/`.

## Original motion brief

User explicitly chose continuing Three.js, not AI-generated video. Extend the existing local product studio with a real articulated character, fixed bottle geometry and a deterministic 20-second action sequence.

Delivery promise: motion-led character action, not a still image or billboard. English frame text, no audio. Chinese interface and provenance.

Character source: the three supplied photos of the red-shirt/yellow-shorts vinyl Shin-chan figurine, including its thick brows, closed eyes, O mouth and exaggerated head/cheeks. Geometry is a reference-inspired approximation, not an official/scanned character model.

Product source: prior original photograph and two videos. Reuse original-photo markings. Use an open mouth because the supplied long video starts open; park the cap on the tabletop. Do not mime drinking through a closed cap.

Action: anticipate, reach, grip, lift, align rim to lips, tilt/sip, lower, set down, release, react. Grip targets are bottle-local anchors; wrist/elbow positions are solved from those anchors. During drinking a selected rim contact point follows the character's mouth anchor. The cap stays on the table and its tether updates with the moving bottle.

Scope/fidelity: body proportions adjusted from visual comparison only; no exact physical dimensions. Fixed geometry provides temporal identity, not proven physical fidelity. Soft cord movement is a procedural visual curve, not a physical simulation.

Deliverables: shinchan.html interactive scene, time scrubbing, orbit/zoom, camera presets, geometry view, 20-second portrait MP4, source/reference panel, QA reports.

Validation: inspect silhouette and face from references; inspect reach/grip, rim/lip contact, cap state, set-down, no obvious prop/head penetration; record anchor errors throughout held/drinking intervals; confirm deterministic seeking, mobile layout, playback and media specs.