Production interior authoring and interactive Three.js delivery plugin for img2.
Oak / Linen: lessons transferred into this workflow
This is one authorized synthetic-reference living room, not a benchmark of automatic reconstruction. The source project is kiss-gs-real-photo-experiment; its ledger is notes/img2interior-distillation.md and its acceptance receipts are under evidence/interior/. Those external assets/receipts are not bundled here. Treat the figures below as historical case evidence, not universal tolerances or fresh proof for a new run.
| Observed failure or boundary | Required action now |
|---|---|
| Technical gates passed while table/chair/pillow forms were rejected | Review camera and component form before finishes; keep rejection evidence |
| EXR orientation wrong without the one required row reversal | Calibrate a rotated asymmetric UV1 fixture against the actual renderer/decoder |
| Cloth point mismatch and initially expensive solver | Sample render-lattice/material coordinates, preserve UV0, measure held error and settle time |
| Hidden/offscreen state, capture loss and wheel conflicts | Own input only during grip; cancel on interruption; measure sleep after settling |
| Controls updates below EPS skipped change events | Clamp eye/target after every update, test real keyboard/orbit/pan/zoom, restore wrappers on disposal |
| Curtain closure could be faked by global dimming | Compare fixed-camera receiver/control patches; separate ambient bake from actual direct shadows |
| Resource dictionary did not attach albedo | Inspect actual exported/runtime texture bindings, not asset acquisition success |
| Roughness image overrode a matte scalar | Distinguish scalar checks from decoded texture/effective response and macro visual review |
| Blender 4.3 sheen export lost constant weight | Verify non-white exported tint; scope conversion to export and restore native inputs in finally |
| Packed image reuse kept old PNG pixels | Refresh owned packed image, reopen without the original file and compare pixels |
| Capture beyond viewport changed canvas width/aspect | Use in-viewport capture; record bounds/projection before relaxing an input invariant |
| Imported quaternion camera ignored Euler-only changes | Set rotation mode deliberately, inspect world forward/target alignment and actual footprint |
| Portable baked UV1 material blurred in macro views | Label representation limits; reopen native/runtime/portable separately |
| Two public GLBs were mistaken for two initial loads | Inspect actual requests and optional downloads; editable master is not another GLB |
Historical current-case observations
- 208 mesh objects, 661,342 triangles, 18 components; dimensions/unseen surfaces remain authored estimates.
- Final cloth sampled grip errors: approximately 0.016 mm near and 0.932 mm wide; zero camera-component delta in those exercises, preserved UV0 and rail headers.
- Cloth wind-off rendering settled to the same frame count over the observed interval. Context loss/retry and three sofa finishes were exercised; this does not prove real-phone behavior.
- Final runtime/portable Khronos receipts reported 0 errors/0 warnings; that is structural validity, not visual acceptance.
- Runtime GLB 106,821,020 bytes; portable 44,913,856 bytes; native master 85,250,515 bytes. Full viewer assets were about 409.50 MB before JS/CSS, a research payload rather than a lightweight deployment target.
- Static indirect light/probe remain based on open curtains. The portable is a static appearance snapshot, not exported browser cloth physics or close-up PBR equivalence.
Do not promote case settings to recovered measurements
The 0.65 m curtain tile, high-roughness dye, normal strength 0.70, low sheen, shadow transmission approximation and room bounds are authored settings. They do not measure fibre size, BRDF, absolute image scale or physically exact transport. The workflow promotes causal checks and evidence discipline, not fixed numerical styling for every interior.
Promotion boundary
No old global skill files or stopped Jev worktrees are changed by this package. The old procedural Forge scaffold is not represented as a completed Blender pipeline. New projects run the installed plugin's actual tools and author their own evidence; historical receipts do not grant new approvals.