MotionEdit: Benchmarking and Learning Motion-Centric Image Editing

✨ Overview

MotionEdit is a novel dataset and benchmark for motion-centric image editing. We also propose MotionNFT (Motion-guided Negative-aware FineTuning), a post-training framework with motion alignment rewards to guide models on motion image editing task.

Model Description

• Model type: Image Editing
• Language(s): English
• Finetuned from model [optional]: Qwen/Qwen-Image-Edit-2509

Model Sources [optional]

• Repository: https://github.com/elainew728/motion-edit/tree/main
• Paper: https://arxiv.org/abs/2512.10284
• Demo Page: https://motion-edit.github.io/

🔧 Usage

🧱 To Start: Environment Setup

Clone our github repository and switch to the directory.

git clone https://github.com/elainew728/motion-edit.git
cd motion-edit

Create and activate the conda environment with dependencies that supports inference and training.

• Note: some models like UltraEdit requires specific dependencies on the diffusers library. Please refer to their official repository to resolve dependencies before running inference.

conda env create -f environment.yml
conda activate motionedit

Finally, configure your own huggingface token to access restricted models by modifying YOUR_HF_TOKEN_HERE in inference/run_image_editing.py.

🔍 Inferencing on MotionEdit-Bench with Image Editing Models

We have released our MotionEdit-Bench on Huggingface. In this Github Repository, we provide code that supports easy inference across open-source Image Editing models: Qwen-Image-Edit, Flux.1 Kontext [Dev], InstructPix2Pix, HQ-Edit, Step1X-Edit, UltraEdit, MagicBrush, and AnyEdit.

Step 1: Data Preparation

The inference script default to using our MotionEdit-Bench, which will download the dataset from Huggingface. You can specify a cache_dir for storing the cached data.

Additionally, you can construct your own dataset for inference. Please organize all input images into a folder INPUT_FOLDER and create a metadata.jsonl in the same directory. The metadata.jsonl file must at least contain entries with 2 entries:

{
    "file_name": IMAGE_NAME.EXT,
    "prompt": PROMPT
}

Then, load your dataset by:

from datasets import load_dataset
dataset = load_dataset("imagefolder", data_dir=INPUT_FOLDER)

Step 2: Running Inference

Use the following command to run inference on MotionEdit-Bench with our MotionNFT Huggingface checkpoint, trained on MotionEdit with Qwen-Image-Edit as the base model:

python inference/run_image_editing.py \
    -o "./outputs/" \
    -m "motionedit" \
    --seed 42

✏️ Citing

@misc{wan2025motioneditbenchmarkinglearningmotioncentric,
      title={MotionEdit: Benchmarking and Learning Motion-Centric Image Editing}, 
      author={Yixin Wan and Lei Ke and Wenhao Yu and Kai-Wei Chang and Dong Yu},
      year={2025},
      eprint={2512.10284},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.10284}, 
}