AI BriefingHunyuan3D ported to Apple Silicon
·2026.04.23 22:55
Key point
Ported Hunyuan3D's image-to-mesh/texture pipeline to Apple Silicon.
Details
Ported Tencent's Hunyuan3D-Paint (texture) and Hunyuan3D-Shape (mesh) to run natively on Apple Silicon.
Replaced CUDA-based nvdiffrast, sparse conv, BVH solver, and CPU unwrapping with Metal GPU kernels, using MLX for texture generation and MPS for mesh generation.
On an M4 Max 40c, the full image → textured mesh pipeline took 3-10 minutes, using about 36GB RAM.
In texture generation benchmarks, MLX was about 4x faster than MPS, with half the memory usage.
- Paint 2.0 (MLX): 114.3s
- Paint 2.0-turbo (MLX): 62.6s
- Paint 2.0 (MPS): 302.4s
- Paint 2.0-turbo (MPS): 222.1s
ComfyUI nodes and MLX weights have been released, while the MLX port of Shape and 2.1 Paint are still WIP.
