FlowOpt

Project | Arxiv | Demo

Official PyTorch implementation of the paper: "FlowOpt: Fast Optimization Through Whole Flow Processes For Training-Free Editing"

Installation

1. Clone the repo

git clone https://github.com/orronai/FlowOpt.git
cd FlowOpt

2. Install the required environment

python -m pip install -r requirements.txt

Usage Example

A. Inversion

Replace <MODEL> with either FLUX or SD3:

python FlowOpt.py --exp_yaml yaml_files/<MODEL>_inversion_exp.yaml

B. Editing

Replace <MODEL> with either FLUX or SD3:

python FlowOpt.py --exp_yaml yaml_files/<MODEL>_editing_exp.yaml

Usage — Your Own Examples

• Upload images to dataset 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 yaml_files/edits.yaml for example.

• Create an experiment file containing the hyperparamaters needed for running FlowOpt, such as T_steps, n_max, max_iterations and eta. For an editing experiment, set the variable inversion to false and for inversion experiment set it to true. This file also includes the path to the yaml_files/edits.yaml file.
  See yaml_files/<MODEL>_editing_exp.yaml for FLUX|SD3 editing usage example, and yaml_files/<MODEL>_inversion_exp.yaml for FLUX|SD3 inversion usage example.
  

  • The provided hyperparameters has been evaluated for images of dimension 1024x1024.

• Run FlowOpt:

python FlowOpt.py --exp_yaml <path_to_your_experiment_yaml>

Gradio GUI

For easy image editing you can also use gradio GUI.

• Install gradio library:

pip install gradio

• Run FlowOpt gradio (examples are included in the demo):

python FlowOpt_gradio.py

License

This project is licensed under the MIT License.

Citation

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

@article{ronai2025flowopt,
	title={FlowOpt: Fast Optimization Through Whole Flow Processes for Training-Free Editing},
	author={Ronai, Or and Kulikov, Vladimir, and Michaeli, Tomer},
	journal={arXiv preprint arXiv:2510.22010},
	year={2025}
}