PhysX-Anything: Simulation-Ready Physical 3D Assets from Single Image

Ziang Cao¹, Fangzhou Hong¹, Zhaoxi Chen¹, Liang Pan², Ziwei Liu¹

¹S-Lab, Nanyang Technological University ²Shanghai AI Laboratory

## 🏆 News - We release the code of PhysX-Anything and our new dataset PhysX-Mobility 🎉 ## PhysX-Anything ### Installation 1. Clone the repo: ``` git clone --recurse-submodules https://github.com/ziangcao0312/PhysX-Anything.git cd PhysX-Anything ``` 2. Create a new conda environment named `physx-anything` and install the dependencies: ```bash . ./setup.sh --new-env --basic --xformers --flash-attn --diffoctreerast --spconv --mipgaussian --kaolin --nvdiffrast ``` **Note**: The detailed usage of `setup.sh` can be found at [TRELLIS](https://github.com/microsoft/TRELLIS) 3. Install the dependencies for Qwen2.5: ```bash pip install transformers==4.50.0 pip install qwen-vl-utils pip install 'accelerate>=0.26.0' ``` **Note**: We release the `requirements.txt` file. You can install all dependencies by running: ```bash conda create -n physx-anything python=3.10 conda activate physx-anything pip install -r requirements.txt ``` ### Inference 1. Download the pre-train model from [huggingface_v1](https://huggingface.co/Caoza/PhysX-Anything). ```bash python download.py ``` 2. Run the inference code ```bash python 1_vlm_demo.py # vlm inference --demo_path ./demo # inputted image path --save_part_ply True # save the geometry of parts --remove_bg False # Set this to false for RGBA images and true otherwise. --ckpt ./pretrain/vlm # ckpt path python 2_decoder.py # decoder inference python 3_split.py # split the mesh python 4_simready_gen.py # convert to URDF & XML --voxel_define 32 # voxel resolution --basepath ./test_demo # results path --process 0 # use postprocess --fixed_base 0 # fix the basement of object or not --deformable 0 # introduce deformable parts or not ``` **Note**: Although our method can generate parts with physical deformable parameters, the deformable components are not stable in MuJoCo. Therefore, we recommend setting the deformable flag to 0 to obtain more reliable simulation results. ### Evaluation 1. Render the generated URDF files ```bash python render_urdf.py ``` 2. Run the VLM-based evaluations. ```bash python evaluation_kine.py ``` **Note**: For all other physical attributes, PhysX-Anything adopts the same settings as [PhysX-3D](https://github.com/ziangcao0312/PhysX-3D). ## PhysX-Mobility For more details about our proposed dataset including dataset structure and annotation, please see this [PhysX-Mobility](https://huggingface.co/datasets/Caoza/PhysX-Mobility) and [PhysXNet](https://huggingface.co/datasets/Caoza/PhysX-3D). ## References If you find PhysX-Anything and PhysX-3D useful for your work, please cite: ``` @article{physxanything, title={PhysX-Anything: Simulation-Ready Physical 3D Assets from Single Image}, author={Cao, Ziang and Hong, Fangzhou and Chen, Zhaoxi and Pan, Liang and Liu, Ziwei}, journal={arXiv preprint arXiv:2511.13648}, year={2025} } @article{physx3d, title={PhysX-3D: Physical-Grounded 3D Asset Generation}, author={Cao, Ziang and Chen, Zhaoxi and Pan, Liang and Liu, Ziwei}, journal={arXiv preprint arXiv:2507.12465}, year={2025} } ``` ### Acknowledgement The data and code is based on [PartNet-mobility](https://sapien.ucsd.edu/browse), [Qwen](https://github.com/QwenLM/Qwen3-VL) and [TRELLIS](https://github.com/microsoft/TRELLIS). We would like to express our sincere thanks to the contributors. ## :newspaper_roll: License Distributed under the S-Lab License. See `LICENSE` for more information.