FlowEdit

Project | Arxiv | Proceedings | Demo | ComfyUI | Data

Recorded Talk

[ICCV 2025 Best Student Paper] Official Pytorch implementation of the paper: "FlowEdit: Inversion-Free Text-Based Editing Using Pre-Trained Flow Models"

Installation

1. Clone the repository

2. Install the required dependencies using pip install torch diffusers transformers accelerate sentencepiece protobuf

   • New version of diffusers may have compatibility issues, try install diffusers==0.30.1
   • Tested with CUDA version 12.4 and diffusers 0.30.0

Running examples

Run editing with Stable Diffusion 3: python run_script.py --exp_yaml SD3_exp.yaml

Run editing with Flux: python run_script.py --exp_yaml FLUX_exp.yaml

Usage - your own examples

• Upload images to example_images folder.

• Create an edits file that specifies: (a) a path to the input image, (b) a source prompt, (c) target prompts, and (d) target codes. The target codes summarize the changes between the source and target prompts and will appear in the output filename.

See edits.yaml for example.

• Create an experiment file containing the hyperparamaters needed for running FlowEdit, such as n_max, n_min. This file also includes the path to the edits.yaml file

See FLUX_exp.yaml for FLUX usage example and SD3_exp.yaml for Stable Diffusion 3 usage example.

For a detailed discussion on the impact of different hyperparameters and the values we used, please refer to our paper.

Run python run_script.py --exp_yaml

ComfyUI implementation for different models

• FLUX
• HunyuanLoom

Implemented by logtd

LTX-Video ComfyUI implementation can be found in LTX-Video official repository.

Community and Follow-Up Work

• Training-Free-WAN-Editing🤗, combines WAN2.1 with FlowEdit to extend training-free to video editing. If you are interested in video editing, please feel free to take a look. Implemented by Kyujinpy.

• DNAEdit refines the Gaussian noise in the noise domain, improving image and video editing (NeurIPS 2025 Spotlight). Project | Code | Arxiv | Proceedings

• FlowAlign add optimal control-based trajectory control to the inversion free process (ICLR 2026). Code | Arxiv | Proceedings

• DynaEdit extened FlowEdit for dynmaic video editing (EECV 2026). Project | Arxiv

License

This project is licensed under the MIT License.

Citation

If you use this code for your research, please cite our paper:

@inproceedings{kulikov2025flowedit,
  title={Flowedit: Inversion-free text-based editing using pre-trained flow models},
  author={Kulikov, Vladimir and Kleiner, Matan and Huberman-Spiegelglas, Inbar and Michaeli, Tomer},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={19721--19730},
  year={2025}
}