TRELLIS.2 image-to-3D hits Apple Silicon

This port is a significant win for the local AI community, proving that even compute-heavy 3D vision models can be decoupled from the CUDA ecosystem. It replaces specialized CUDA kernels like flex_gemm and flash_attn with native MPS-compatible alternatives, achieving generation times around 3.5 minutes on M4 Pro. This enables high-fidelity 3D asset creation with zero cloud cost and full offline privacy. While the sparse 3D convolution implementation is roughly 10x slower than CUDA, it remains highly usable for creative workflows. Technical hurdles like texture baking and spatial hashing were solved via Python-based spatial hashing and torch.grid_sample. The tool requires 24GB+ of unified memory, making it a "Pro" level tool for modern Mac hardware.