Monorepo for Aesthetic.Computer aesthetic.computer
core thespianjas SCORE.md
2.6 kB

thespianjas #

thespianjas is the global Jeffrey digital-twin subsystem for the AC monorepo. It is not owned by one podcast or campaign. Podcast, Reel, lecture, piece, and live-performance pipelines consume versioned assets from here.

Architecture #

  1. Identity — canonical studio references resolve from papers/jeffrey-platter through identity.json.
  2. Reconstruction — provider adapters turn those references into a textured, relightable GLB. Meshy 6 multi-image is the first adapter; Tripo H3.1 and SAM 3D Body are named benchmark candidates.
  3. Studio — metal/ is the primary native macOS renderer. Its SceneKit scene graph is explicitly backed by Metal; key, fill, rim, ground, camera, skeletal playback, and later facial controls remain ours to direct. The browser studio is only a quick compatibility viewer.
  4. Performance — podcast audio drives energy, breathing, head emphasis, and eventually facial blendshapes. A video lipsync provider may polish exported shots, but it is downstream of the canonical relightable twin.
  5. Compositing — marketing/essay-reels adds word-timed captions and social safe-area layout to a rendered thespianjas performance.

The canonical slug is globally unique: thespianjas. Generated binaries are large and versioned locally under assets/versions/vNNN/; manifests and recipes are tracked in git.

First build #

node thespianjas/bin/generate.mjs --version v001 --provider meshy
cd thespianjas/metal && swift run ThespianJas ../assets/versions/v001/idle.glb

Meshy accepts one to four views and returns a textured PBR GLB. The initial studio platter has a strong face view and two seated body views; a dedicated front/side/back neutral-pose capture will materially improve the next rig.

The generated rigged.glb and idle.glb contain the humanoid skeleton. They do not promise facial ARKit blendshapes; viseme/blendshape authoring is a separate performance stage.

Provider notes (July 2026) #

  • Meshy 6 multi-image: 1–4 views, PBR textures, A/T pose options, humanoid auto-rig, GLB/FBX/USDZ/OBJ. Best first complete asset lane.
  • Tripo H3.1: useful single-image benchmark for likeness/geometry.
  • SAM 3D Body: inexpensive human-specific reconstruction benchmark.
  • Kling / Sync lipsync: video finishing layers, not substitutes for the relightable GLB. They accept rendered video or a portrait plus audio.

Identity outputs are publishable only in Jeffrey-operated AC surfaces. Never silently substitute another person's photographs or train from private imagery.