Skip to content

bake-export.md

Latest commit

 

History

History
71 lines (45 loc) · 7.65 KB

bake-export.md

File metadata and controls

71 lines (45 loc) · 7.65 KB
Raw

Calibrate → bake → probe → export

Prerequisites

An actually rendered, editable native scene with finite physical meshes, semantic components, explicit metric UV0, source camera, authored world/lights and source-image aspect. Review camera/form/materials first. Use the pinned bpy environment and an explicit workspace/spec; imports never generate files in the installed plugin.

Coordinate/color calibration, not a realism grade

"$BLENDER_PYTHON" "$SKILL_DIR/tools/author_scene.py" calibrate --workspace "$PWD" --run-id calibration-r01 --width 256 --samples 8
"$BLENDER_PYTHON" "$SKILL_DIR/tools/author_scene.py" material-regression --workspace "$PWD" --run-id packing-r01

calibrate renders and reimports an asymmetric UV1 chart, checks actual glTF texture coordinates/scene bounds/decoded pixels, and renders a real equirectangular EXR with emissive cardinal panels to verify probe orientation. Its runtimeLightmap section emits an additional child viewer workspace, actual asymmetric float EXR field, rotated UV1 PBR model and Cycles expected image. The runtime has zero specular, black probe/background, no sun and PI lightmap intensity; native expected evaluates albedo × field.

Register that child with interior_project.py start --workspace CHILD --image EXPECTED_RENDER --source-kind render, install the real viewer there, npm ci/build/serve, select source camera and capture the actual canvas at the same projection/aspect. Verify its drawing-buffer dimensions and real device pixel ratio, not just CSS width; compare against runtimeLightmap.expectedRender without image realignment. This exercises the real EXR row decoder, channel 1 and PI convention; native generation alone does not certify browser sampling. Keep global and clearly defined interior-region errors, capture settings and fixture limits. Do not alter row orientation merely because a room render looks dark.

material-regression uses real PNG/native/GLB files: save packed pixels, change external albedo, refresh, remove external source, reopen and compare actual pixels/UV0; export non-white weighted sheen and verify native restoration on successful and actual filesystem-failed exports. It also repeats atlas preparation and real GLB export, checking stable policy-addressable material names, retained artist roughness and unchanged UV0. It qualifies the exercised Blender 4.3 workaround, not untested exporters or material appearance.

Production bake

"$BLENDER_PYTHON" "$SKILL_DIR/tools/author_scene.py" bake --workspace "$PWD" --run-id final-r01 --resolution 2048 --samples 64

The pipeline groups physical surfaces into architecture/textiles/furniture/garden/details, skips incompatible glass/bulbs, keeps UV0 and creates a separate Lightmap UV1. It checks finite/noncollapsed UVs and allocates chart gutters. Textile/garden atlases use twice the requested base resolution; choose a budget from actual footprint, not an unlabelled universal quality claim.

Subsequent bakes reuse prepared materials and their existing UV0 mapping nodes rather than growing material-name suffixes or copies. Preserve the reserved InteriorBakeTarget ownership node when editing prepared native materials; material policies address the actual stable exported names.

Profile a labelled low-resolution mechanism smoke before the final bake. A tiny mesh collapsing at a small atlas is a real UV-budget failure: increase that group's resolution; never remove the mesh, relax the pixel-area guard or present the smoke as final quality. --atlas GROUP saves that completed group's receipt and packed master independently, allowing different justified per-group budgets. Full 2K/4K CPU bakes can outlive a command runner's deadline; use an appropriate long-running job, retain receipts and inspect its actual completion before probing/exporting. A timed-out bake is not success.

For each group it actually performs:

  1. Full diffuse direct+indirect lighting without albedo color, linear EXR evidence.
  2. Object-normal-guided HDR OpenImageDenoise filtering of lighting only.
  3. UV1 emission bake of original UV0 albedo, optional alpha, and full-light × albedo static portable PNG with the chosen display transform once.
  4. A second diffuse pass with direct SUN disabled, then restore native SUN energy in finally.
  5. Denoised linear SUN-free ambient EXR for runtime, with PI-compensated lightmap intensity in the manifest.

Bake receipts include groups/objects, resolutions/samples, actual statistics, denoiser, lighting model and elapsed time. OIDN failure is a real failure; do not substitute a fake noise-free atlas. UV1 baking must not move source sampling away from UV0.

--atlas textiles scopes a repair; retained groups preserve old transport approximations. A globally changed layout/light/reflectance requires full regeneration. Do not call a partial bake a globally current physical GI solution.

Packed image repair and probe freshness

"$BLENDER_PYTHON" "$SKILL_DIR/tools/author_scene.py" refresh-material --workspace "$PWD" --run-id material-r02 --image public/assets/interior/curtain-basecolor.png --image public/assets/interior/curtain-roughness.png
"$BLENDER_PYTHON" "$SKILL_DIR/tools/author_scene.py" bake --workspace "$PWD" --run-id material-r02 --resolution 2048 --samples 64
"$BLENDER_PYTHON" "$SKILL_DIR/tools/author_scene.py" probe --workspace "$PWD" --run-id material-r02 --width 1024 --samples 96

Use actual named used maps, not showcase filenames for another scene. Refresh unpacks only selected stale images, reloads disk pixels, repacks and preserves topology/UV; it invalidates atlases and the authoring revision. Draft also starts a new revision. Probe renders the current scene at the declared indoor pose with the measured native/glTF orientation and records revision/pose/actual EXR hash. Export rejects an old/mutated probe even when a file of that name exists. Out-of-band Blender edits still require explicit dependent regeneration; this is not a sandbox or universal scene-diff engine.

Export distinct products

"$BLENDER_PYTHON" "$SKILL_DIR/tools/author_scene.py" export --workspace "$PWD" --run-id final-r01

Export checks complete current bake inventory/model and clean required atlases. It reconstructs authoring receipt from the reopened master rather than requiring a prior run's evidence folder. Source dimensions, view names, room bounds, probe/background, optional sun, finishes and credit derive from this scene/spec.

  • room-master.blend: actual packed editable scene, geometry, UV0/UV1, PBR and lights. This is not a GLB.
  • room-runtime.glb: real PBR with UV0/UV1 and semantic/interaction extras; EXR atlases/probe and optional HDR background remain required sidecars.
  • room-portable.glb: independent embedded full-static diffuse appearance, KHR_materials_unlit sampling UV1; not a browser shader/physics/dye-state export.
  • manifest.json: schema-2 runtime paths/camera/bounds/lightmaps/lighting/variants/limits and receipt-derived live API credit.

Export-only tangent-cap triangulation, native thin-pane output substitution and constant weighted-sheen tint conversion are temporary and restored even on failure. Portable material swaps also restore native slots in finally. Native GI/probe remain static as cloth/finishes change; glass transmission is a declared approximation, not real-time caustics.

Exit criteria

Open the source/lateral/material/contact renders and actual viewer. Run current material/delivery audits, Khronos validation and independent inspect_scene.py master|portable; compare source and close views separately. Keep reports/hashes/versioned evidence, packing/color-transform facts and static/runtime limits. A zero exit code, triangle count or calibration MAE is not visual acceptance.