Mesh Generation

1. There is no standalone "mesh generation" model — meshes are the OUTPUT of text-to-3D and image-to-3D pipelines. The setup is the same as those tasks.
2. For clean topology meshes (not point clouds, not Gaussian splats): Hunyuan3D-2 produces the best open-weight meshes. pip install gradio + clone the repo.
3. The pipeline: input (text or image) → multi-view diffusion → mesh reconstruction (Marching Cubes or similar isosurface extraction) → mesh refinement (decimation, smoothing) → UV unwrapping → texture baking.
4. Hunyuan3D-2 outputs .glb files with polygon meshes. Import into Blender → inspect topology → retopologize if needed.
5. For CAD-quality meshes (parametric, NURBS): no open-weight AI model produces these yet. AI generates polygon meshes; CAD conversion is a separate manual step.
6. First usable mesh in 2-10 minutes on 12+ GB GPU. For simple shapes: Shap-E gives meshes in 30-60 seconds on CPU.

Used RTX 3060 12 GB (~$200-250, see /hardware/rtx-3060-12gb). Runs Hunyuan3D-2 at 5-15 minutes for textured meshes. TripoSR at 10-30 seconds for simple shapes. For indie game development (props, furniture, basic environment pieces): $400 hardware generates usable meshes after manual cleanup. The bottleneck is NOT GPU speed — it's the manual retopology and cleanup. AI mesh generation on $400 hardware is viable for prototyping; production game assets need significant human post-processing regardless of GPU tier.

Used RTX 3090 24 GB (~$700-900, see /hardware/rtx-3090). Runs Hunyuan3D-2 at 2-5 minutes for high-quality meshes with detailed textures. For a professional 3D pipeline (game studio, archviz, VFX): one RTX 3090 generates 50-100 mesh variations/day for art director review. Total: ~$1,800-2,200. Mesh generation quality plateaus above 12 GB — the model architecture, not hardware, limits output quality. A 3090 vs. 4090 changes generation speed by 50%, not mesh quality. Invest in 3D art skills before investing in a 4090 for mesh gen.

The mistake: Exporting an AI-generated mesh directly into a game engine and wondering why the collision detection doesn't work, the lighting looks wrong, and the LOD (level of detail) system can't generate lower-poly versions. Why it fails: AI meshes are "dirty" — inconsistent triangle density (dense in detailed areas, sparse in flat areas), non-manifold edges, no LOD groups, no collision primitives, UV seams in visible areas, inverted normal faces. Game engines expect clean, optimized meshes. The fix: Import into Blender or similar DCC tool. Steps always needed: (1) Decimate to target triangle count, (2) Recalculate normals, (3) Fix non-manifold geometry (3D Print Toolbox → Make Manifold), (4) Manually adjust UV seams, (5) Create LODs (Decimate modifier with different ratios), (6) Add simple collision mesh. AI generates the rough sculpt; you produce the game-ready asset. Budget 20-60 minutes of manual work per AI-generated mesh for game production.