memgrafter/conjure3d-mac-gguf

Conjure3D for Apple Silicon: TRELLIS.2 and BiRefNet in GGUF

These are the weights for the experimental Apple Silicon route of Conjure3D. The route runs trellis.cpp (a C++/GGML port of TRELLIS.2) on Metal.

Status: experimental. The output is not the same as the CUDA route. Use it to try images and seeds on a Mac. Make the assets that you ship on the CUDA route.

Files

File	Source (pinned)
ss_flow.gguf	TRELLIS.2-4B@af44b45f ckpts/ss_flow_img_dit_1_3B_64_bf16.safetensors
ss_dec.gguf	TRELLIS.2-4B@af44b45f microsoft/TRELLIS-image-large/ckpts/ss_dec_conv3d_16l8_fp16.safetensors
shape_flow_512.gguf	TRELLIS.2-4B@af44b45f ckpts/slat_flow_img2shape_dit_1_3B_512_bf16.safetensors
shape_flow_1024.gguf	TRELLIS.2-4B@af44b45f ckpts/slat_flow_img2shape_dit_1_3B_1024_bf16.safetensors
shape_dec.gguf	TRELLIS.2-4B@af44b45f ckpts/shape_dec_next_dc_f16c32_fp16.safetensors
tex_flow_512.gguf	TRELLIS.2-4B@af44b45f ckpts/slat_flow_imgshape2tex_dit_1_3B_512_bf16.safetensors
tex_flow_1024.gguf	TRELLIS.2-4B@af44b45f ckpts/slat_flow_imgshape2tex_dit_1_3B_1024_bf16.safetensors
tex_dec.gguf	TRELLIS.2-4B@af44b45f ckpts/tex_dec_next_dc_f16c32_fp16.safetensors
birefnet.gguf	ZhengPeng7/BiRefNet@e2bf8e44 model.safetensors
manifest.json	The SHA-256 of each GGUF file, its source files and their pins

The route also needs dinov3.gguf (the image features that TRELLIS.2 uses). It has a different licence and is in a gated repository: memgrafter/conjure3d-mac-dinov3-gguf. You can also make it yourself with mac/convert_weights.py after you accept the DINOv3 License.

How the files were made

mac/convert_weights.py in Conjure3D does these steps:

1. It checks the SHA-256 of each source file against its pin.
2. It runs the converter of trellis.cpp (tools/convert.py, trellis.cpp c0bed38c, tag v0.8.1). Matrices are stored as f16 and 1-D tensors as f32.
3. It writes manifest.json.

To check a download, compare the SHA-256 of each file with manifest.json.

Use

Set up Conjure3D as docs/MAC.md tells you (config.env, mac/setup_mac.sh build and env), then use "Weights, way A" there. In short:

1. Open memgrafter/conjure3d-mac-dinov3-gguf with your own account, read the DINOv3 License and accept it if you agree.

2. Download both repositories into the same folder, $TRELLIS_GGUF:

   $MAC_VENV/bin/hf auth login
   $MAC_VENV/bin/hf download memgrafter/conjure3d-mac-gguf --local-dir "$TRELLIS_GGUF" --exclude birefnet.gguf
   $MAC_VENV/bin/hf download memgrafter/conjure3d-mac-dinov3-gguf --local-dir "$TRELLIS_GGUF"
   

3. The cut-out step uses the BiRefNet safetensors file, not birefnet.gguf. Download it at the pinned revision into $HF_HOME:

   $MAC_VENV/bin/hf download ZhengPeng7/BiRefNet model.safetensors README.md --revision e2bf8e4460fc8fa32bba5ea4d94b3233d367b0e4
   

Each run checks every GGUF file that it loads against manifest.json and stops if one is missing or different.

• birefnet.gguf is the BiRefNet conversion that mac/convert_weights.py also makes. The Conjure3D route does not use it.
• Needs: macOS on Apple Silicon. On an Apple M6 with 24 GiB, one object took 202-803 s (median 312 s) with the draft profile, and peak memory was 2.65-5.30 GiB (measured 2026-11-01).
• Disk: 15.0 GB for the files that the route uses (15.4 GB with birefnet.gguf), 0.6 GB for dinov3.gguf.

Licence

MIT. TRELLIS.2 (code and weights) and BiRefNet (code and weights) are MIT. These files are format conversions of their weights.