The execution model is in blender-core. Always re-run blender-asset-qa after optimizing.
scripts/mesh_stats.py -- --objects ... --budget mobile|roblox|web|vr|console|pc. It returns tris, estimated draw calls, materials, texture VRAM, bones, and the over-budget items.make_lods.py --cleanup (merge by distance 0.1 mm, dissolve degenerate, delete loose). Remove hidden faces: back-to-back faces and faces under other parts (qa_check → back_to_back, plus manual xray).blender-texture-baking).scripts/make_lods.py -- --objects SM_X --rename creates _LOD1.._LOD3 at 50/25/12.5%. Check every LOD silhouette with look/review_shots.blender-export-pipelines).| Platform | Hero char tris | Prop tris | Mats/asset | Max texture | Bones |
|---|---|---|---|---|---|
| Mobile | 5–15k | 0.3–3k | 1–2 | 1024 | ≤75 |
| Roblox | body ≤10,742 (head 4k, torso 1.75k, limb 1,248); accessory/clothing 4k | ≤20k per MeshPart | 1 (SurfaceAppearance) | 1024 (PBR), 4k max | R15 + optional |
| Web (three.js/Babylon) | 10–30k | 1–10k | 1–2 | 2048 | ≤100 |
| Standalone VR (Quest) | 10–40k | 1–10k | 1–3 | 2048 | ≤100 |
| Console / PC | 50–150k | 5–50k | 2–6 | 4096 | 150–300 |
| UE5 Nanite (static) | — | millions OK | few | 4096+ | n/a |
Whole-scene limits matter more than single assets: mobile ≈ 100–300k tris visible and < 100–200 draw calls; PC ≈ millions of tris and < 2–5k draw calls.
| Mode | Use | Notes |
|---|---|---|
COLLAPSE (ratio) | organic, scans, LODs | use_symmetry, vertex group factor to protect faces/hands; triangulates |
UNSUBDIV (iterations) | meshes made by subdividing quads | restores the quad grid cleanly |
DISSOLVE (planar, angle) | hard-surface / CAD | delimit {"SEAM","SHARP","MATERIAL","UV"}; keeps the silhouette, makes n-gons → triangulate on export |
Protect UV seams and silhouette: decimate a copy, then compare. If UVs tear at LOD1, the ratio is too aggressive for that topology, so re-unwrap or use a manual LOD.
GPU vertex count = unique (position, normal, UV, color) combinations. Every UV seam and hard edge duplicates vertices, so engine vertex counts are often 1.5–3× Blender's. Minimize seams where they aren't needed, keep hard edges = seams (so the split happens once), and don't add vertex colors or extra UV sets you don't use.
lod_bias auto).RenderFidelity=Automatic does it. Keep each MeshPart ≤ 20k tris and use a simple CollisionFidelity (Box/Hull) for non-walkable parts.EXT_mesh_gpu_instancing with export_gpu_instances) or as one mesh + transforms in the engine.Keep collision separate and simple: boxes/convex hulls of ≤ 32–255 verts. UE UCX_<Mesh>_##, Godot -colonly/-convcolonly suffix, Unity MeshCollider on a simplified mesh, Roblox CollisionFidelity. Never use the render mesh as a physics mesh for anything that moves.
Limit to 4 influences per vertex (vertex_group_limit_total(limit=4) + normalize). Remove non-deforming bones from the export (use_armature_deform_only). Bake constraints and drivers to keys. Shape keys cost memory per vertex, so drop unused ones and keep only deltas (glTF sparse accessors handle this).
make_lods.py evaluates them first. Other paths need you to apply them.Ctrl+J) merges materials and UV maps by name. Rename the UV maps identically first, or the UVs get lost.Weld modifier is the non-destructive merge by distance. Use mode="CONNECTED" to avoid welding separate parts.