TRELLIS-image-large for diffusers-3d
microsoft/TRELLIS-image-large converted into a
diffusers-3d pipeline: ordinary Diffusers component folders
(config.json + safetensors) plus the object3d_model_index.json sidecar that the package's auto-loader
validates before downloading anything. Nothing here requires remote code.
Install
pip install git+https://github.com/suvadityamuk/diffusers.git
pip install "git+https://github.com/suvadityamuk/diffusers.git#subdirectory=packages/diffusers-3d"
diffusers-3d runs every network in plain PyTorch on CPU or GPU. Rendering Gaussian splats needs the optional
gsplat backend; meshing and PBR export for TRELLIS.2 need the compiled O-Voxel runtime (see the package docs).
Usage
import torch
from diffusers_3d import AutoPipelineForImageTo3D, ImageCondition
pipeline = AutoPipelineForImageTo3D.from_pretrained("suvadityamuk/TRELLIS-image-large-diffusers-3d", dtype=torch.bfloat16).to("cuda")
output = pipeline(ImageCondition(image=rgba), formats=("gaussian", "mesh", "radiance_field"))
gaussians, mesh, radiance_field = output.objects
rgba is a (4, H, W) tensor in [0, 1] whose alpha channel masks the object; the pipeline crops and recentres
it as the released code does. formats selects any of sparse_structure, slat, gaussian, mesh, and
radiance_field.
Components
| Folder | Class | Released file |
|---|---|---|
conditioner | TrellisDinov2Conditioner | facebook/dinov2-with-registers-large (DINOv2 ViT-L/14 with registers) |
sparse_structure_flow_model | TrellisSparseStructureFlowModel | ss_flow_img_dit_L_16l8_fp16 |
sparse_structure_decoder | TrellisSparseStructureDecoder | ss_dec_conv3d_16l8_fp16 |
slat_flow_model | TrellisSLatFlowModel | slat_flow_img_dit_L_64l8p2_fp16 |
gaussian_decoder | TrellisSLatGaussianDecoder | slat_dec_gs_swin8_B_64l8gs32_fp16 |
mesh_decoder | TrellisSLatMeshDecoder | slat_dec_mesh_swin8_B_64l8m256c_fp16 |
radiance_field_decoder | TrellisSLatRadianceFieldDecoder | slat_dec_rf_swin8_B_64l8r16_fp16 |
Both schedulers carry the released sampler settings (25 steps, guidance 5.0 over the 0.5–1.0 interval,
rescale_t=3). Weights are stored as released (float16 for the transformers); load with dtype= to pick the
compute precision.
Provenance
Converted with diffusers-3d-convert-trellis from diffusers-3d 0.1.0.dev0 against TRELLIS revision
442aa1e1afb9014e80681d3bf604e8d728a86ee7. The conversion renames parameters into the package layout and reformats configs; it does not
change any weight value. Tiny-configuration parity tests against the pinned upstream code are part of the package
test suite.
License and attribution
TRELLIS weights and architecture: MIT License, Copyright (c) Microsoft Corporation. The DINOv2 conditioner weights are Apache-2.0, Copyright (c) Meta Platforms, Inc. This repository redistributes both under those terms; it is not affiliated with or endorsed by Microsoft or Meta.