Show HN: Running TRELLIS.2 Image-to-3D Generation Natively on Apple Silicon

Ported Microsoft TRELLIS.2's image-to-3D model to run on Apple Silicon using PyTorch MPS.

• Runs natively on Mac without an NVIDIA GPU
• On M4 Pro, 24GB, generates a 400K+ vertex mesh from a single image in about 3.5 minutes
• Output is saved as textured OBJ and GLB, including PBR materials, ready to use directly in 3D apps

The porting work includes:

• pure-PyTorch / gather-scatter-based backends/conv_none.py instead of flex_gemm
• Python dictionary-based backends/mesh_extract.py instead of o_voxel._C hashmap
• PyTorch SDPA applied instead of flash_attn
• cumesh and nvdiffrast handled via stubs or graceful skip
• Hardcoded .cuda() calls throughout the code replaced based on the active device

Performance was measured with pipeline-type 512.

• model loading: ~45s
• image preprocessing: ~5s
• sparse structure sampling: ~15s
• shape SLat sampling: ~90s
• texture SLat sampling: ~50s
• mesh decoding: ~30s
• total: ~3.5 min

The limitations are also clear.

• Texture export not possible: without nvdiffrast, only vertex color is output
• Hole filling disabled: small holes may remain since cumesh is not supported
• About 10x slower than CUDA: pure-PyTorch sparse convolution is the bottleneck
• Training not supported, inference only