The pipeline.json in microsoft/TRELLIS.2-4B references:
"sparse_structure_decoder": "microsoft/TRELLIS-image-large/ckpts/ss_dec_conv3d_16l8_fp16"
This checkpoint is from the original TRELLIS repo and is not compatible with TRELLIS.2's flow model. When used, the decoder produces all-negative values (range approximately -200 to -70) regardless of input, resulting in zero valid voxels and broken 3D reconstruction.
Expected: TRELLIS.2 should include its own sparse_structure_decoder checkpoint trained to match its flow model output distribution.
Workaround attempted: Using a percentile threshold (top 50%) instead of > 0 produces voxels but they are geometrically incorrect (cube-shaped noise).
Environment: Google Colab A100, PyTorch 2.10, CUDA 12.8
Please either include the correct TRELLIS.2 sparse structure decoder in the model release, or update pipeline.json to point to a compatible checkpoint.
The
pipeline.jsoninmicrosoft/TRELLIS.2-4Breferences:"sparse_structure_decoder": "microsoft/TRELLIS-image-large/ckpts/ss_dec_conv3d_16l8_fp16"
This checkpoint is from the original TRELLIS repo and is not compatible with TRELLIS.2's flow model. When used, the decoder produces all-negative values (range approximately -200 to -70) regardless of input, resulting in zero valid voxels and broken 3D reconstruction.
Expected: TRELLIS.2 should include its own
sparse_structure_decodercheckpoint trained to match its flow model output distribution.Workaround attempted: Using a percentile threshold (top 50%) instead of
> 0produces voxels but they are geometrically incorrect (cube-shaped noise).Environment: Google Colab A100, PyTorch 2.10, CUDA 12.8
Please either include the correct TRELLIS.2 sparse structure decoder in the model release, or update
pipeline.jsonto point to a compatible checkpoint.