The execution model is in blender-core. Material wiring and measured values are in blender-pbr-materials, and bakes are in blender-texture-baking. This skill authors the surface story: where paint chips, where dirt collects, what's printed on it.
scripts/smart_material.py -- --objects Crate --wear 0.5 --grime 0.4 --dust 0.3
scripts/smart_material.py -- --objects Crate --rust 0.6 --under-color 0.3 0.29 0.28 # rusty iron
scripts/smart_material.py -- --objects SM_Crate --wear 0.5 --curvature-map //tex/T_Crate_CURV.png --ao-map //tex/T_Crate_AO.png
scripts/decal.py -- --target Wall --at 0 -0.1 1.5 --image //decals/hazard.png --size 0.4
scripts/smart_material.py -- --objects Crate --remove # exact restore
Start with a clean base (correct albedo/roughness, blender-pbr-materials), then stack from the inside out: base → edge wear → rust → grime → dust. Each layer has a physical cause:
| Layer | Mask (cause) | Look |
|---|---|---|
| Edge wear | convex curvature (edges get hit) + chip noise | reveals the under-material: bare metal, primer, raw wood |
| Rust | cavities and worn edges + blotch noise | orange-brown, metallic 0, roughness ~0.85 |
| Grime | concave curvature + (1 − AO) (dirt collects where it can't be wiped) | dark, rougher, not metallic |
| Dust | world up-facing normal × patch noise (it settles from above) | light, roughness ~0.9 |
| Breakup | fine noise on roughness (±25%) | the cheapest realism win; never ship a flat roughness value |
Pure curvature looks procedural and even. Always multiply in noise so wear is heavy in some places and absent in others. Less is more: wear 0.2–0.4 reads as "used", and 0.7+ reads as "abandoned".
TX_Layers group in front of every Principled BSDF (Base Color, Roughness, Metallic) and records the originals, so --remove restores links and values exactly. Re-running updates values in place. All layer amounts stay inputs on the group node, so you can tweak them there.TX_Curvature point attribute is the mean elevation of the neighbours within --edge-width (m, default 0.02). It's density-independent: band ≈ edge width. Set the width to roughly the bevel width (1–3 cm on props, 5–10 cm on buildings). Cost: ~1.5 s per 30k verts.note: wear would smear across whole faces. On low-poly game meshes, bake from the high (bake_maps.py --high ... --maps normal ao curvature) and pass --curvature-map/--ao-map (loaded as Non-Color, read through --uv).--scale = feature size in m), so masks stick to the asset when it moves. Mirrored/linked duplicates share the pattern, so vary --scale or the object if that shows.--under-color 0.45 0.32 0.2 --under-metallic 0 --under-roughness 0.7.Engines don't run Blender node trees. Bake the result into a texture set: bake_maps.py -- --low SM_Crate --maps basecolor roughness metallic normal --pack-orm (self-bake, no --high). Then build_pbr_material.py --textures and export. Check the result with blender-asset-qa and a lit look.
--at, oriented to the normal and spun by --rotation. A Shrinkwrap (project, above surface) keeps it --offset (2 mm) off the surface. It's parented, has no shadow, uses Dithered alpha, and its height follows the image aspect.blender-asset-qa does, and doubles --subdiv (to 128) until flat triangles stop dipping over convex creases. The result reports gap_min_m/gap_max_m. Never place decals at 0 offset: that's guaranteed z-fighting.Decal) for dynamic ones, or bake them into the base texture.Trim sheets, tiling with vertex-color blends, hand/texture painting (5.2 brush assets) and headless pixel writes: references/trims-tiling-painting.md.
--scale. Both are world-sized.TX_Curvature is mesh data. Applying modifiers or remeshing afterward invalidates it, so re-run after topology changes.