UniPhysGen-1.7B-Physics
UniPhysGen-1.7B-Physics is the physical semantic alignment checkpoint of UniPhysGen, with direct support for part-level intrinsic physical grounding. It combines the Qwen3-1.7B language backbone with the Sonata point-cloud encoder and a lightweight MLP projector.
Given a full-object point cloud and a target-part point cloud, it predicts structured part identity, semantic descriptions, and intrinsic physical properties. During alignment pretraining, this supervision connects local part geometry and global object context with functional semantics, articulation behavior, and intrinsic physical properties, establishing a shared physical grounding representation.
The checkpoint therefore serves both as a part-level physics predictor and as the shared initialization for downstream task-specific fine-tuning of UniPhysGen-1.7B-Kinematics, UniPhysGen-1.7B-Structure, and UniPhysGen-1.7B-Object. See Section 4.1 and Appendices B.1 and C of the paper for the alignment objective and training stages.
Model details
| Item | Value |
|---|---|
| Checkpoint role | Physical semantic alignment; part-level physics inference; downstream fine-tuning initialization |
| Internal task name | physics |
| Required geometry | Object point cloud + target-part point cloud |
| Training lineage | UniPhysGen-1.7B-Init → physical semantic alignment pretraining → this checkpoint |
| Training data | breezexian/UniPhys-40K |
| Evaluation data | UniPhys-Bench · UniPhys-Bench Part 2 |
| Tested Transformers version | 4.51.0 |
| Source code | breezexian/UniPhysGen |
| Paper | arXiv:2607.13586 |
Inputs and outputs
The recommended point-cloud format is .npz with aligned arrays:
point float32 [N, 3]
color uint8 [N, 3]
normal float32 [N, 3]
Object and part point clouds are normalized together by the inference pipeline. The structured result contains:
part_identity: part name and motion type;semantic_description: basic, functional, movement, and grasp descriptions;physical_properties: material, density, Young's modulus, hardness, Poisson's ratio, friction, graspability, and affordance.
Installation
The model has been tested on Linux with Python 3.11, PyTorch 2.4.1, CUDA 12.4,
and transformers==4.51.0.
[!IMPORTANT] Use
transformers==4.51.0. This is the tested version and is pinned by the UniPhysGen project metadata.
git clone https://github.com/breezexian/UniPhysGen.git
cd UniPhysGen/uniphysgen
conda create -n uniphysgen python=3.11 -y
conda activate uniphysgen
conda install -y -c nvidia/label/cuda-12.4.0 cuda-toolkit
python -m pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu124
python -m pip install -e ".[train]"
bash scripts/install_cuda_extensions.sh
For inference only, replace python -m pip install -e ".[train]" with
python -m pip install -e ..
The model uses a custom Transformers architecture. Install UniPhysGen before
loading the checkpoint; a generic transformers.pipeline is not supported.
Inference
CUDA_VISIBLE_DEVICES=0 python inference_batch_intrinsic_physics_part.py \
--model_path breezexian/UniPhysGen-1.7B-Physics \
--object_pcd examples/object.npz \
--part_pcd examples/part.npz \
--output outputs/intrinsic_physics_part.json
The entry point also accepts a JSON list through --input_json. See the main
project README for the batch schema, path resolution rules, and output record
format. The final prediction is stored in result, while raw_response keeps
the direct generated text.
Evaluation
Evaluate predictions using the UniPhysGen evaluator:
python -m eval intrinsic_physics_part PREDICTIONS \
--output intrinsic_physics_part_metrics.json
See Table 2 of the paper for the UniPhys-Bench results and the main project README for the complete evaluation protocol.
Intended use and limitations
This checkpoint is intended for research on physical grounding of 3D object parts, initialization of downstream grounding models, and producing candidate simulation parameters that are subsequently validated. It is not a physics simulator, and its estimates should not be treated as measured material properties.
Performance may degrade for incomplete scans, sparse or noisy point clouds, unseen materials, ambiguous part decompositions, unusual scales, or objects far outside the training distribution. Validate predictions before using them in robotics, safety-critical systems, or real-world engineering workflows.
License
The model weights are released under the Creative Commons Attribution-NonCommercial 4.0 International license. Commercial use is not permitted under this license. The UniPhysGen source code is licensed separately under Apache-2.0. See the included license for Qwen3 and Sonata attribution.
Citation
@article{li2026uniphysgen,
title = {UniPhysGen: Unified Physical Grounding for Simulation-Ready 3D Assets},
author = {Li, Xian and Wei, Rong and Yang, Lujie and Huang, Haolin and Fang, Junyuan and Tang, Siliang and Xiao, Jun and Tang, Rui and Li, Juncheng},
journal = {arXiv preprint arXiv:2607.13586},
year = {2026}
}