- Category
- AI Tools
- Rank
- No. 337Tools index
- Pricing
- Open Source
- Type
- TOOL
- Builder
- microsoft
- GitHub
- 13.6k stars
- Date
About
Microsoft Research's structured 3D latent diffusion model for scalable 3D asset generation from text or images. CVPR 2025 Spotlight.
What it does
TRELLIS builds an internal structured representation of an object, then decodes it into radiance fields, 3D Gaussians, or meshes. It accepts text, one image, or multiple views. Generated results can be rendered to video, converted into textured GLB meshes, or saved as Gaussian PLY files. It also supports creating variants from an existing mesh.
Why it's ranked here
The strongest case is breadth within one generation system. The same pipeline covers image conditioning, text conditioning, multiple views, asset variants, three internal output representations, and practical exports. Pretrained models and browser demos lower the experimentation barrier. The research depth is substantial, but the demanding installation and GPU requirements keep it firmly in specialist territory.
What's good
Output choice is unusually practical. A single generation can provide Gaussians for splatting, radiance fields for rendering, and meshes for conventional 3D workflows. The supplied postprocessing turns Gaussian and mesh outputs into textured GLB assets with configurable simplification and texture size. Training code and dataset preparation tools also make the repository useful beyond inference.
Tradeoffs
The documented setup needs Linux, an NVIDIA GPU with at least 16GB of memory, CUDA compilation, and a large dependency stack. Windows guidance is not fully tested. The authors recommend image conditioning because text models are less creative and detailed. Multi-image conditioning uses a tuning-free method and may struggle with inconsistent details or poses.
How to use it well
Use TRELLIS for research, technical prototyping, or batch creation where you control a capable Linux GPU environment. For text ideas, first create a strong reference image, then run the image-conditioned model and export GLB or PLY as needed. Use multiple views only when they show consistent geometry and appearance. It does not cover the upstream text-to-image step that its recommended workflow depends on.
Technical notes+
README.md documents Conda-oriented installation through setup.sh, selectable flash-attn or xformers attention, pretrained Hugging Face checkpoints, and Python 3.8 or newer. example.py and example_text.py load TrellisImageTo3DPipeline and TrellisTextTo3DPipeline, move them to CUDA, render all three representations, and export GLB and PLY. app.py adds single-image and experimental multi-image Gradio flows, while app_text.py provides a separate text demo. train.py exposes JSON-configured, retrying multi-GPU and multi-node training through argparse and torch.multiprocessing. dataset_toolkits/render.py, dataset_toolkits/download.py, dataset_toolkits/voxelize.py, dataset_toolkits/render_cond.py, and dataset_toolkits/stat_latent.py form a command-line data preparation surface using CSV metadata, Blender rendering, voxelization, and latent statistics.
Observed
- Primary language
- Python
- Install surface
- Repository clone with submodules, Conda-oriented shell setup, and selectable dependency groups
- Interfaces
- Python pipelines, Gradio web demos, training CLI, and dataset preparation CLIs
- Platform support
- Tested only on Linux; Windows setup is documented as not fully tested
- Hardware requirement
- NVIDIA GPU with at least 16GB memory and a CUDA Toolkit
- Output formats
- Radiance fields, 3D Gaussians, meshes, textured GLB, PLY, and rendered video
Read from README.md, app.py, train.py, example.py, app_text.py, example_text.py, example_variant.py, example_multi_image.py, trellis/__init__.py, dataset_toolkits/utils.py, dataset_toolkits/render.py, dataset_toolkits/download.py, dataset_toolkits/voxelize.py, dataset_toolkits/render_cond.py, dataset_toolkits/stat_latent.py.
What it can do
Generate 3D models from text descriptions
Text prompt describing desired 3D object → 3D asset/model
Generate 3D models from images
Input image → 3D asset/model
Create structured 3D representations using latent diffusion
Text or image input → Structured 3D latent representation
Scale 3D asset generation for batch processing
Multiple text prompts or images → Multiple 3D assets
Convert 2D visual concepts to 3D objects
2D image or visual description → 3D geometric model
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.