Key takeaways
- Separate network transfer from VRAM: compressed image bytes govern HTTP download time, whereas a single decoded 1024 × 1024 RGBA8 texture consumes ~4 MiB of base GPU memory (~5.33 MiB with mipmaps).
- Baseline every source file before editing: record byte size, triangle count, vertex count, material count (draw calls), texture pixel dimensions, and Y-up metre bounding boxes.
- Stop simplifying before silhouette, UV borders, or doorway openings break: our Shipping Crate stays at 12 triangles in both tiers because there is nothing meaningful to decimate.
- Use explicit tier ceilings: SceneCrate’s automated static-prop pipeline enforces <= 15,000 triangles / 1024 px / 4 materials for Standard and <= 5,000 triangles / 512 px / 2 materials for Lightweight.
1. Record the original baseline before changing anything
Keep an untouched copy of the source GLB and record its exact byte length, triangle count, vertex count, material count, texture resolutions, and physical dimensions in metres. Count instanced placements alongside unique meshes: one 5,000-triangle prop drawn 80 times without instancing or material batching still multiplies draw calls and vertex work.
Always separate HTTP download size from GPU memory allocation. JPEG, PNG, or WebP compression inside a GLB shrinks network transfer, but standard browser texture decoding expands a 1024 × 1024 RGBA8 image to roughly 4 MiB for its base mip level alone, and a 4096 × 4096 map to 64 MiB before mipmaps.
- Confirm the GLB embeds all buffers and textures with zero external URI dependencies.
- Audit unsupported or costly features such as unused animations, morph targets, or alpha-blended sorting.
- Pick budgets for your lowest target mobile or integrated-GPU browser, not a desktop workstation.
2. Simplify geometry and maps only where quality holds
Decimate dense curvature while locking UV seam borders, hard normal edges, ground contact points, and architectural openings. Preserve the exact metre height and Y-up orientation across LOD tiers so swapping Standard for Lightweight never shifts gameplay alignment.
Resize base-color maps separately from tangent-space normal and packed metallic-roughness maps. Normal and roughness channels store linear directional or surface data rather than photographs; aggressive lossy compression or careless color-space conversion introduces visible shading bands.
SceneCrate’s automatic convex-prop checks enforce 15,000 / 5,000 triangle limits, 1024 / 512 px texture caps, and 4 / 2 material limits across Standard and Lightweight. Complex multi-material scenes require tailored rules: Mossgate Passage uses 2 baked PBR atlases across 105,945 triangles (Standard) and 103,781 triangles (Lightweight), locking UV borders while halving atlas dimensions from 4K to 2K.
3. Reload the exported GLB and verify real metrics
Open the final exported GLB in your actual Three.js scene rather than trusting DCC viewport previews. Check double-sided vs backface-culled triangles, ground alignment at y = 0, and collider fit next to a 1.7 m player reference. Measure frame time, draw calls, and GPU memory separately from zip size.
Read actual output bytes instead of assuming smaller textures always compress smaller. Across our free prop cases, Moss Stone drops from 1.31 MiB to 320.3 KiB at 80 triangles, whereas Oak Barrel goes from 113.5 KiB (Standard) to 207.7 KiB (Lightweight) at 1,244 triangles, and Shipping Crate goes from 54.2 KiB to 118.1 KiB at 12 triangles.
Frequently asked questions
Why does a 2 MB GLB file still use tens of megabytes of browser GPU memory?
Compressed PNG or JPEG textures inside a GLB only stay compressed over the network. When WebGL uploads standard 2D textures, each pixel expands to 4 bytes of RGBA8 data plus roughly 33% extra for mipmaps, so four 2048 × 2048 maps take over 85 MiB of VRAM.
What is a practical triangle budget for static browser-game props?
For individual solid props in WebGL, up to 15,000 triangles with 1024 px textures works well for close-up Standard tiers, while 500 to 5,000 triangles with 512 px textures suits Lightweight or mobile tiers. Simple boxes like crates only need 12 triangles.
Why should UV borders be locked when simplifying baked PBR models?
When textures are baked from a high-poly sculpt onto UV islands, moving vertices along UV seams stretches the atlas across gutter boundaries and creates dark or jagged seam lines.
Does reducing material count matter as much as reducing triangles?
Often more on web browsers. In Three.js, each distinct material on a mesh primitive submits a separate WebGL draw call, so merging props onto 1 to 4 shared materials frequently saves more CPU frame time than shaving a few hundred triangles.