TRELLIS-text-large for diffusers-3d
microsoft/TRELLIS-text-large converted into a diffusers-3d pipeline. The text release shares its decoders with the image release; they are included here so the repository loads on its own.
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 AutoPipelineForTextTo3D
pipeline = AutoPipelineForTextTo3D.from_pretrained("suvadityamuk/TRELLIS-text-large-diffusers-3d", dtype=torch.bfloat16).to("cuda")
output = pipeline("a wooden rocking chair", formats=("gaussian", "mesh"))
Prompts may also be TextCondition(text=..., negative_text=...) values. Defaults follow the released text
sampler (guidance 7.5 over the 0.5–0.95 interval).
Components
| Folder | Class | Released file |
|---|---|---|
conditioner | TrellisClipTextConditioner | openai/clip-vit-large-patch14 text tower and tokenizer |
sparse_structure_flow_model | TrellisSparseStructureFlowModel | ss_flow_txt_dit_L_16l8_fp16 |
sparse_structure_decoder | TrellisSparseStructureDecoder | ss_dec_conv3d_16l8_fp16 (from TRELLIS-image-large) |
slat_flow_model | TrellisSLatFlowModel | slat_flow_txt_dit_L_64l8p2_fp16 |
gaussian_decoder | TrellisSLatGaussianDecoder | slat_dec_gs_swin8_B_64l8gs32_fp16 (from TRELLIS-image-large) |
mesh_decoder | TrellisSLatMeshDecoder | slat_dec_mesh_swin8_B_64l8m256c_fp16 (from TRELLIS-image-large) |
radiance_field_decoder | TrellisSLatRadianceFieldDecoder | slat_dec_rf_swin8_B_64l8r16_fp16 (from TRELLIS-image-large) |
Provenance
Converted with diffusers-3d-convert-trellis from diffusers-3d 0.1.0.dev0 against TRELLIS revision
442aa1e1afb9014e80681d3bf604e8d728a86ee7. Weight values are unchanged.
License and attribution
TRELLIS weights and architecture: MIT License, Copyright (c) Microsoft Corporation. The CLIP text encoder weights are MIT, Copyright (c) OpenAI. Not affiliated with or endorsed by Microsoft or OpenAI.