UniPhysGen-1.7B-Init
UniPhysGen-1.7B-Init is the training initialization checkpoint for UniPhysGen, a unified 3D model for grounding articulation semantics and intrinsic physical properties. It combines the Qwen3-1.7B language backbone with the Sonata point-cloud encoder and the UniPhysGen multimodal projection layers.
This checkpoint contains no UniPhys task-specific grounding training. It is intended as the starting point for physical semantic alignment and subsequent task-specific fine-tuning. It is not a recommended inference checkpoint.
Model details
| Item | Value |
|---|---|
| Model architecture | Qwen3-1.7B + Sonata + UniPhysGen multimodal projector |
| Intended stage | Training initialization |
| Language backbone | Qwen/Qwen3-1.7B |
| Point-cloud backbone | facebook/sonata |
| Input modality | 3D point clouds, with task-dependent text prompts |
| Tested Transformers version | 4.51.0 |
| Source code | breezexian/UniPhysGen |
| Paper | arXiv:2607.13586 |
Intended use
Use this checkpoint to reproduce the physical semantic alignment stage or to train UniPhysGen on compatible custom data. For released predictions, use one of the task-specific checkpoints instead:
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 the UniPhysGen package before loading the checkpoint so that the custom configuration and model classes are registered.
Download and training
hf download breezexian/UniPhysGen-1.7B-Init \
--local-dir checkpoints/UniPhysGen-1.7B-Init
Set the initialization path in the semantic-alignment release config:
model_name_or_path: checkpoints/UniPhysGen-1.7B-Init
Then launch the training entry point documented in the main project README. The corresponding training dataset is breezexian/UniPhys-40K.
Limitations
- This checkpoint is an initialization artifact and has not been fine-tuned for any of the four released grounding tasks.
- It should not be used as a drop-in text-generation model or with a generic Hugging Face text-generation pipeline.
- Inputs must follow the UniPhysGen point-cloud and prompt conventions.
- Model outputs may be inaccurate and require task-specific fine-tuning and validation before use in simulation or robotics.
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.
This checkpoint is derived from Qwen3-1.7B, released under Apache-2.0, and the official Sonata pretrained weights, released under CC BY-NC 4.0. See the included license for attribution details.
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}
}