PhyEdit

This repository contains the release LoRA checkpoint for PhyEdit: Towards Real-World Object Manipulation via Physically-Grounded Image Editing.

• Base model: Qwen/Qwen-Image-Edit-2511
• Training data: RealManip-40K
• Code and ManipEval: nenhang/PhyEdit

Files

File	Description
phyedit_lora.safetensors	PhyEdit LoRA weights
training_config.json	Sanitized release training configuration
SHA256SUMS	Weight checksum

Usage

Download the checkpoint:

hf download ruihangxu/PhyEdit phyedit_lora.safetensors \
  --local-dir checkpoints/PhyEdit

Clone the code repository and run ManipEval sampling:

git clone https://github.com/nenhang/PhyEdit.git
cd PhyEdit

CUDA_VISIBLE_DEVICES=0 \
python -m bench.sample \
  --config-path configs/train_deepspeed.yaml \
  --pretrained-model-path Qwen/Qwen-Image-Edit-2511 \
  --checkpoint-path ../checkpoints/PhyEdit/phyedit_lora.safetensors \
  --benchmark-metadata data/RealManip-40K/metadata/test.json \
  --output-dir outputs/manipeval \
  --base-size 1024 \
  --batch-size 8 \
  --seeds 42 43 44 45 46 47 48 49

The release checkpoint was trained with an aspect-ratio-preserving base area of 589824 (768^2). This corresponds to 1024 x 576 for 16:9 images and 768 x 768 for square images.

Limitations

This is a LoRA adapter and requires the Qwen-Image-Edit-2511 base model. Image quality and geometric accuracy depend on the source image, masks, depth and camera estimates, movement magnitude, and sampling seed. Review the licenses and usage terms of the base model, dataset, and external geometry models before use.

License

The PhyEdit LoRA weights are released under the MIT License. Third-party models and datasets retain their own licenses and terms.

Citation

@misc{xu2026phyeditrealworldobjectmanipulation,
      title={PhyEdit: Towards Real-World Object Manipulation via Physically-Grounded Image Editing},
      author={Ruihang Xu and Dewei Zhou and Xiaolong Shen and Fan Ma and Yi Yang},
      year={2026},
      url={https://arxiv.org/abs/2604.07230},
}