research(050): auto-detected procedural relief presets from one photo

[dompm]

What

Makes the glass preview's surface-relief effect (kept as a presentation-layer effect per 047's decision) driven by a preset picked automatically from one user photo, per the CTO's hard requirement (no trying 8 settings).

Delivers:

1. Preset bank (render050/relief_presets.py) — 6 categories × ≤3 knobs (smooth, hammered, granite, seedy, ripple, rolling_wave), procedural, deterministic, tileable (band-limited Fourier noise + wrapped seed-stamps). Taxonomy grounded 1:1 in the generator's own generate_relief_height families + real corpus classes. Board: results/050/board_preset_bank.jpg.
2. Auto-detector (render050/detect_relief.py) — VLM multiple-choice category (claude CLI, haiku, direction-first prompt — the validated classifier pattern) + classical and VLM knob bins (compared) + a T-derived pseudo-normal blend component.
3. Synthetic quantitative validation (gen050_*, score050.py) — front-lit volumetric-slab "user photos" of recipes with known authored relief; scored with holdout discipline (seed 6001 tune / 7001-7002 holdout).
4. Real qualitative board (board_real.py) — corpus/wild sheets + the CTO's difficult photo. results/050/board_real_detection.jpg.
5. End-to-end perceptual A/B (model050.html, drive050_ab.py) — the 047 scene in real three.js MeshPhysicalMaterial: GT normal | procedural preset | off. results/050/board_ab_normal.jpg.

Honest finding (negative-ish, characterized)

From one photo, the coarse relief character is only weakly detectable and the fine sub-label + numeric knobs are not:

• Holdout category: fine ≈ 0.50–0.58, coarse ≈ 0.50–0.67 across two runs (chance 0.17/0.25); VLM is stochastic (~±2/12). Only smooth and ripple are reliable across both runs.
• Not a model-capability ceiling: sonnet is worse (0.33/0.42) → genuine single-photo ambiguity. The generator itself authors hammered/granite/seedy as one shared isotropic base, so they are near-identical surfaces.
• Knobs weak: classical amp 0.42 / scale 0.17; VLM amp 0.50 / scale 0.42.
• A/B is the good news: the bank's preset for the right coarse group ties with GT perceptually (both textured, both ≠ flat "off"), so collapsing the fine trio costs nothing visible.

Recommendation

Ship auto-detection for the axis one photo supports (flat? directional/streaky?), default the rest to a single textured preset (hammered, medium knobs) + T-blend on streaky. Satisfies "no 8 settings" while being honest about the fine limit. Full details in reports/050-auto-relief-presets.md.

Raw renders/assets gitignored (deterministically regenerable); boards + JSON committed. Do not merge — for lab-lead review.

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying vitrai with    Cloudflare Pages

Latest commit:	4c0dd3e
Status:	✅  Deploy successful!
Preview URL:	https://35e1641a.vitrai.pages.dev
Branch Preview URL:	https://research-delighting-050-auto.vitrai.pages.dev

View logs