AI BriefingHunyuan3D 3x Faster
·2026.04.24 02:26
Key point
Hunyuan3D-Paint and Shape were ported to MLX/MPS for Apple Silicon, making texture generation 3x faster.
Details
Tencent Hunyuan3D-Paint (texture generation) and Hunyuan3D-Shape (mesh generation) were each ported to MLX and MPS respectively to run on Apple Silicon. CUDA-based nvdiffrast, sparse conv, BVH solver, and CPU unwrapping were replaced with Metal GPU kernels.
- MLX texture generation is about 3x faster than MPS, with half the memory usage.
- On an M4 Max 40c, the image-to-textured-mesh pipeline takes 3-10 minutes depending on model choice, with RAM usage of about 36GB.
- Moving Shape generation to MLX as well is still WIP, so there's room for further improvement.
- ComfyUI nodes and MLX weights have already been released.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Underdog Releases 'Husky', a Model-Specific Inference Engine Up to 4.5x Faster than MLX
Qwen3.8-27B-Uncensored-MLX (4 min read)