metadata
license: mit
pipeline_tag: image-to-image
tags:
- physics-aware
- image-editing
PhysicEdit: Physics-Aware Image Editing with Latent Transition Priors
PhysicEdit is an end-to-end framework for physics-aware image editing that reformulates the editing process as predictive physical state transitions. By combining a frozen Qwen2.5-VL for physically grounded reasoning with learnable transition queries, it provides visual guidance to a diffusion backbone to ensure results are both semantically aligned and physically plausible (e.g., handling refraction or material deformation).
- Paper: From Statics to Dynamics: Physics-Aware Image Editing with Latent Transition Priors
- Project Page: https://liangbingzhao.github.io/statics2dynamics/
- Repository: https://github.com/liangbingzhao/PhysicEdit
Usage
To use this model, please follow the installation instructions in the official repository. You can perform inference on a single image using the provided validation script:
python scripts/inference/validate.py \
--prompt "your_editing_instruction" \
--image_path /path/to/input.jpg \
--save_path /path/to/output.jpg \
--base_model_path /path/to/Qwen-Image-Edit-2509 \
--dinov2_path /path/to/DINOv2-with-registers-base \
--lora_path /path/to/physicedit_checkpoint.safetensors
Citation
If you find this work useful for your research, please cite the following BibTeX:
@misc{zhao2026staticsdynamicsphysicsawareimage,
title={From Statics to Dynamics: Physics-Aware Image Editing with Latent Transition Priors},
author={Liangbing Zhao and Le Zhuo and Sayak Paul and Hongsheng Li and Mohamed Elhoseiny},
year={2026},
eprint={2602.21778},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2602.21778},
}