FreeFlux: Understanding and Exploiting Layer-Specific Roles in RoPE-Based MMDiT for Versatile Image Editing (ICCV2025)

Tianyi Wei   Yifan Zhou   Dongdong Chen   Xingang Pan

S-lab, Nanyang Technological University; Microsoft GenAI

This repository hosts the official PyTorch implementation of the paper: "FreeFlux: Understanding and Exploiting Layer-Specific Roles in RoPE-Based MMDiT for Versatile Image Editing".

Leveraging the layer-specific roles in RoPE-based MMDiT we discovered, versatile training-free image editing is tailored to different task characteristics, including non-rigid editing, object addition, background replacement, object movement, and outpainting.

Getting Started

Prerequisites

$ conda create -n freeflux python=3.10
$ pip install torch==2.5.1+cu121 torchvision==0.20.1+cu121 --index-url https://download.pytorch.org/whl/cu121
$ pip install diffusers==0.31.0 transformers==4.46.2
$ pip install matplotlib jupyter ipykernel opencv-python scipy sentencepiece protobuf accelerate
$ git clone https://github.com/facebookresearch/sam2.git && cd sam2
$ pip install -e .

Probing

$ cd probing
$ python3 gen_rope_probing_dataset.py
$ python3 positional_dependency_eval.py

To probe the layer-wise positional dependency in FLUX, navigate to the probing directory and first run gen_rope_probing_dataset.py to generate the probing dataset. Then, execute positional_dependency_eval.py to evaluate the dependency levels across different layers.

Enjoy

$ cd editing

To start the editing experience for a specific task, simply open the .ipynb file located in the directory corresponding to the task name.

If you run it on a remote server, you need to run jupyter notebook --port=20367 --allow-root --ip=0.0.0.0 first. then use e.g. VS Code to select that Jupyter Server as your kernel.

Citation

If you find our work useful for your research, please consider citing the following papers :)

@article{wei2025freeflux,
  title={FreeFlux: Understanding and Exploiting Layer-Specific Roles in RoPE-Based MMDiT for Versatile Image Editing},
  author={Wei, Tianyi and Zhou, Yifan and Chen, Dongdong and Pan, Xingang},
  journal={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  year={2025}
}