TRELLIS is a large 3D asset generation model. It takes in text or image prompts and generates high-quality 3D assets in various formats, such as Radiance Fields, 3D Gaussians, and meshes. The cornerstone of TRELLIS is a unified Structured LATent (SLAT) representation that allows decoding to different output formats and Rectified Flow Transformers tailored for SLAT as the powerful backbones. We provide large-scale pre-trained models with up to 2 billion parameters on a large 3D asset dataset of 500K diverse objects. TRELLIS significantly surpasses existing methods, including recent ones at similar scales, and showcases flexible output format selection and local 3D editing capabilities which were not offered by previous models.
Check out our Project Page for more videos and interactive demos!
- High Quality: It produces diverse 3D assets at high quality with intricate shape and texture details.
- Versatility: It takes text or image prompts and can generate various final 3D representations including but not limited to Radiance Fields, 3D Gaussians, and meshes, accommodating diverse downstream requirements.
- Flexible Editing: It allows for easy editings of generated 3D assets, such as generating variants of the same object or local editing of the 3D asset.
03/25/2025
- Release training code.
- Release TRELLIS-text models and asset variants generation.
- Examples are provided as example_text.py and example_variant.py.
- Gradio demo is provided as app_text.py.
- Note: It is always recommended to do text to 3D generation by first generating images using text-to-image models and then using TRELLIS-image models for 3D generation. Text-conditioned models are less creative and detailed due to data limitations.
12/26/2024
- Release TRELLIS-500K dataset and toolkits for data preparation.
12/18/2024
- Implementation of multi-image conditioning for TRELLIS-image model. (#7). This is based on tuning-free algorithm without training a specialized model, so it may not give the best results for all input images.
- Add Gaussian export in
app.pyandexample.py. (#40)
- System: The code is currently tested only on Linux. For windows setup, you may refer to #3 (not fully tested).
- Hardware: An NVIDIA GPU with at least 16GB of memory is necessary. The code has been verified on NVIDIA A100 and A6000 GPUs.
- Software:
- The CUDA Toolkit is needed to compile certain submodules. The code has been tested with CUDA versions 11.8 and 12.2.
- Conda is recommended for managing dependencies.
- Python version 3.8 or higher is required.
Clone the repo:
git clone --recurse-submodules https://github.com/microsoft/TRELLIS.git cd TRELLISInstall the dependencies:
Before running the following command there are somethings to note:
- By adding
--new-env, a new conda environment namedtrelliswill be created. If you want to use an existing conda environment, please remove this flag. - By default the
trellisenvironment will use pytorch 2.4.0 with CUDA 11.8. If you want to use a different version of CUDA (e.g., if you have CUDA Toolkit 12.2 installed and do not want to install another 11.8 version for submodule compilation), you can remove the--new-envflag and manually install the required dependencies. Refer to PyTorch for the installation command. - If you have multiple CUDA Toolkit versions installed,
PATHshould be set to the correct version before running the command. For example, if you have CUDA Toolkit 11.8 and 12.2 installed, you should runexport PATH=/usr/local/cuda-11.8/bin:$PATHbefore running the command. - By default, the code uses the
flash-attnbackend for attention. For GPUs do not supportflash-attn(e.g., NVIDIA V100), you can remove the--flash-attnflag to installxformersonly and set theATTN_BACKENDenvironment variable toxformersbefore running the code. See the Minimal Example for more details. - The installation may take a while due to the large number of dependencies. Please be patient. If you encounter any issues, you can try to install the dependencies one by one, specifying one flag at a time.
- If you encounter any issues during the installation, feel free to open an issue or contact us.
Create a new conda environment named
trellisand install the dependencies:. ./setup.sh --new-env --basic --xformers --flash-attn --diffoctreerast --spconv --mipgaussian --kaolin --nvdiffrastThe detailed usage of
setup.shcan be found by running. ./setup.sh --help.Usage: setup.sh [OPTIONS] Options: -h, --help Display this help message --new-env Create a new conda environment --basic Install basic dependencies --train Install training dependencies --xformers Install xformers --flash-attn Install flash-attn --diffoctreerast Install diffoctreerast --spconv Install spconv --mipgaussian Install mip-splatting --kaolin Install kaolin --nvdiffrast Install nvdiffrast --demo Install all dependencies for demo
- By adding
We provide the following pretrained models:
| Model | Description | #Params | Download |
|---|---|---|---|
| TRELLIS-image-large | Large image-to-3D model | 1.2B | Download |
| TRELLIS-text-base | Base text-to-3D model | 342M | Download |
| TRELLIS-text-large | Large text-to-3D model | 1.1B | Download |
| TRELLIS-text-xlarge | Extra-large text-to-3D model | 2.0B | Download |
Note: It is always recommended to use the image conditioned version of the models for better performance.
Note: All VAEs are included in TRELLIS-image-large model repo.
The models are hosted on Hugging Face. You can directly load the models with their repository names in the code:
TrellisImageTo3DPipeline.from_pretrained("microsoft/TRELLIS-image-large")If you prefer loading the model from local, you can download the model files from the links above and load the model with the folder path (folder structure should be maintained):
TrellisImageTo3DPipeline.from_pretrained("/path/to/TRELLIS-image-large")Here is an example of how to use the pretrained models for 3D asset generation.
importos# os.environ['ATTN_BACKEND'] = 'xformers' # Can be 'flash-attn' or 'xformers', default is 'flash-attn'os.environ['SPCONV_ALGO'] ='native'# Can be 'native' or 'auto', default is 'auto'.# 'auto' is faster but will do benchmarking at the beginning.# Recommended to set to 'native' if run only once.importimageiofromPILimportImagefromtrellis.pipelinesimportTrellisImageTo3DPipelinefromtrellis.utilsimportrender_utils, postprocessing_utils# Load a pipeline from a model folder or a Hugging Face model hub.pipeline=TrellisImageTo3DPipeline.from_pretrained("microsoft/TRELLIS-image-large")
pipeline.cuda()
# Load an imageimage=Image.open("assets/example_image/T.png")
# Run the pipelineoutputs=pipeline.run(
image,
seed=1,
# Optional parameters# sparse_structure_sampler_params={# "steps": 12,# "cfg_strength": 7.5,# },# slat_sampler_params={# "steps": 12,# "cfg_strength": 3,# },
)
# outputs is a dictionary containing generated 3D assets in different formats:# - outputs['gaussian']: a list of 3D Gaussians# - outputs['radiance_field']: a list of radiance fields# - outputs['mesh']: a list of meshes# Render the outputsvideo=render_utils.render_video(outputs['gaussian'][0])['color']
imageio.mimsave("sample_gs.mp4", video, fps=30)
video=render_utils.render_video(outputs['radiance_field'][0])['color']
imageio.mimsave("sample_rf.mp4", video, fps=30)
video=render_utils.render_video(outputs['mesh'][0])['normal']
imageio.mimsave("sample_mesh.mp4", video, fps=30)
# GLB files can be extracted from the outputsglb=postprocessing_utils.to_glb(
outputs['gaussian'][0],
outputs['mesh'][0],
# Optional parameterssimplify=0.95, # Ratio of triangles to remove in the simplification processtexture_size=1024, # Size of the texture used for the GLB
)
glb.export("sample.glb")
# Save Gaussians as PLY filesoutputs['gaussian'][0].save_ply("sample.ply")After running the code, you will get the following files:
sample_gs.mp4: a video showing the 3D Gaussian representationsample_rf.mp4: a video showing the Radiance Field representationsample_mesh.mp4: a video showing the mesh representationsample.glb: a GLB file containing the extracted textured meshsample.ply: a PLY file containing the 3D Gaussian representation
app.py provides a simple web demo for 3D asset generation. Since this demo is based on Gradio, additional dependencies are required:
. ./setup.sh --demoAfter installing the dependencies, you can run the demo with the following command:
python app.pyThen, you can access the demo at the address shown in the terminal.
We provide TRELLIS-500K, a large-scale dataset containing 500K 3D assets curated from Objaverse(XL), ABO, 3D-FUTURE, HSSD, and Toys4k, filtered based on aesthetic scores. Please refer to the dataset README for more details.
TRELLIS’s training framework is organized to provide a flexible and modular approach to building and fine-tuning large-scale 3D generation models. The training code is centered around train.py and is structured into several directories to clearly separate dataset handling, model components, training logic, and visualization utilities.
- train.py: Main entry point for training.
- trellis/datasets: Dataset loading and preprocessing.
- trellis/models: Different models and their components.
- trellis/modules: Custom modules for various models.
- trellis/pipelines: Inference pipelines for different models.
- trellis/renderers: Renderers for different 3D representations.
- trellis/representations: Different 3D representations.
- trellis/trainers: Training logic for different models.
- trellis/utils: Utility functions for training and visualization.
Prepare the Environment:
- Ensure all training dependencies are installed.
- Use a Linux system with an NVIDIA GPU (The models are trained on NVIDIA A100 GPUs).
- For distributed training, verify that your nodes can communicate through the designated master address and port.
Dataset Preparation:
- Organize your dataset similar to TRELLIS-500K. Specify your dataset path using the
--data_dirargument when launching training.
- Organize your dataset similar to TRELLIS-500K. Specify your dataset path using the
Configuration Files:
- Training hyperparameters and model architectures are defined in configuration files under the
configs/directory. - Example configuration files include:
- Training hyperparameters and model architectures are defined in configuration files under the
The training script can be run as follows:
usage: train.py [-h] --config CONFIG --output_dir OUTPUT_DIR [--load_dir LOAD_DIR] [--ckpt CKPT] [--data_dir DATA_DIR] [--auto_retry AUTO_RETRY] [--tryrun] [--profile] [--num_nodes NUM_NODES] [--node_rank NODE_RANK] [--num_gpus NUM_GPUS] [--master_addr MASTER_ADDR] [--master_port MASTER_PORT]
options:
-h, --help show this help message and exit
--config CONFIG Experiment config file
--output_dir OUTPUT_DIR Output directory
--load_dir LOAD_DIR Load directory, default to output_dir
--ckpt CKPT Checkpoint step to resume training, default to latest
--data_dir DATA_DIR Data directory
--auto_retry AUTO_RETRY Number of retries on error
--tryrun Try run without training
--profile Profile training
--num_nodes NUM_NODES Number of nodes
--node_rank NODE_RANK Node rank
--num_gpus NUM_GPUS Number of GPUs per node, default to all
--master_addr MASTER_ADDR Master address for distributed training
--master_port MASTER_PORT Port for distributed trainingTo train a image-to-3D stage 2 model with a single machine.
python train.py \
--config configs/vae/slat_vae_dec_mesh_swin8_B_64l8_fp16.json \
--output_dir outputs/slat_vae_dec_mesh_swin8_B_64l8_fp16_1node \
--data_dir /path/to/your/dataset1,/path/to/your/dataset2 \The script will automatically distribute the training across all available GPUs. Specify the number of GPUs with the --num_gpus flag if you want to limit the number of GPUs used.
To train a image-to-3D stage 2 model with multiple GPUs across nodes (e.g., 2 nodes):
python train.py \
--config configs/generation/slat_flow_img_dit_L_64l8p2_fp16.json \
--output_dir outputs/slat_flow_img_dit_L_64l8p2_fp16_2nodes \
--data_dir /path/to/your/dataset1,/path/to/your/dataset2 \
--num_nodes 2 \
--node_rank 0 \
--master_addr $MASTER_ADDR \
--master_port $MASTER_PORTBe sure to adjust node_rank, master_addr, and master_port for each node accordingly.
By default, training will resume from the latest saved checkpoint in the same output directory. To specify a specific checkpoint to resume from, use the --load_dir and --ckpt flags:
python train.py \
--config configs/generation/slat_flow_img_dit_L_64l8p2_fp16.json \
--output_dir outputs/slat_flow_img_dit_L_64l8p2_fp16_resume \
--data_dir /path/to/your/dataset1,/path/to/your/dataset2 \
--load_dir /path/to/your/checkpoint \
--ckpt [step]- Auto Retry: Use the
--auto_retryflag to specify the number of retries in case of intermittent errors. - Dry Run: The
--tryrunflag allows you to check your configuration and environment without launching full training. - Profiling: Enable profiling with the
--profileflag to gain insights into training performance and diagnose bottlenecks.
Adjust the file paths and parameters to match your experimental setup.
TRELLIS models and the majority of the code are licensed under the MIT License. The following submodules may have different licenses:
diffoctreerast: We developed a CUDA-based real-time differentiable octree renderer for rendering radiance fields as part of this project. This renderer is derived from the diff-gaussian-rasterization project and is available under the LICENSE.
Modified Flexicubes: In this project, we used a modified version of Flexicubes to support vertex attributes. This modified version is licensed under the LICENSE.
If you find this work helpful, please consider citing our paper:
@article{xiang2024structured,
title = {Structured 3D Latents for Scalable and Versatile 3D Generation},
author = {Xiang, Jianfeng and Lv, Zelong and Xu, Sicheng and Deng, Yu and Wang, Ruicheng and Zhang, Bowen and Chen, Dong and Tong, Xin and Yang, Jiaolong},
journal = {arXiv preprint arXiv:2412.01506},
year = {2024}
}