SplitFlow (NeurIPS 2025)

SplitFlow: Flow Decomposition for Inversion-Free Text-to-Image Editing

(Sung-Hoon Yoon *, Minghan Li *, Gaspard Beaudouin, Congcong Wen, Muhammad Rafay Azhar, Mengyu Wang)

SplitFlow extends FlowEdit by leveraging LLMs to decompose editing prompt into progressive manner, enabling more natural image editing results even in complex scenarios.

Installation

1. Clone the repository

git clone https://github.com/Harvard-AI-and-Robotics-Lab/SplitFlow
cd SplitFlow

2. Create conda environment and install dependencies

conda create -n splitflow python=3.8
conda activate splitflow

3. Install PyTorch (CUDA 12.1)

pip install torch==2.4.1 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

4. Install required packages

pip install diffusers==0.30.1 transformers==4.46.3 accelerate==1.0.1 \
            sentencepiece==0.2.0 protobuf==5.29.3 matplotlib

5. LLM Model Setup

• The code automatically downloads mistralai/Mistral-7B-Instruct-v0.3 model

Running Examples

Run editing with Stable Diffusion 3:

python run_script.py --exp_yaml SD3_SplitFlow.yaml

Usage - Your Own Examples

1. Prepare Input Images

Upload images you want to edit to the example_images folder.

2. Create Edit Dataset File

Create an edits.yaml file that specifies:

• input_img: Path to the input image
• source_prompt: Prompt describing the source image
• target_prompts: Prompt describing the target edit

Example:

-
    input_img: example_images/your_image.jpg
    source_prompt: A group of sheep stands on a green meadow
    target_prompts:
    - Three dogs with sunglasses stands on a green meadow with snowy mountains

3. Create Experiment Configuration File

Create an experiment file following the format of SD3_SplitFlow.yaml:

• exp_name: Experiment name (output directory name)
• dataset_yaml: Path to the edit dataset file
• model_type: Model type to use (currently supports "SD3")
• T_steps: Number of sampling steps (default: 50)
• n_min, n_max: Timestep range where editing is applied (0~33 recommended)
• src_guidance_scale: Source guidance scale (default: 3.5)
• tar_guidance_scale: Target guidance scale (default: 13.5)
• seed: Random seed

For detailed discussion on hyperparameters, please refer to the FlowEdit,ICCV 2025 paper.

4. Run

python run_script.py --exp_yaml <your_experiment.yaml>

5. Check Results

Edited images are saved at outputs/<exp_name>/<model_type>/src_<source_name>/tar_<target_id>/. Each output folder contains:

• Edited image (PNG file)
• prompts.txt: Prompt and configuration information used

Project Structure

SplitFlow/
├── run_script.py              # Main execution script
├── SplitFlow_utils.py         # SplitFlow algorithm implementation
├── SD3_SplitFlow.yaml         # SD3 experiment configuration file
├── edits.yaml                 # Edit dataset file
├── assets/                    # README images and resources
├── example_images/            # Input images folder
└── outputs/                   # Output results folder

Algorithm Overview

SplitFlow edits images through the following process:

1. Prompt Decomposition: Uses LLM to decompose complex target prompts into 3 intermediate steps
2. Editing with Decomposed Flows: Applies each intermediate prompt sequentially while performing flow matching
3. Aggregation and Final Editing: Applies the final target prompt to generate the final image

This approach divides complex editing tasks into multiple steps, enabling more natural and controllable editing results.

Supported Models

Currently supports the following models:

• Stable Diffusion 3 (SD3): stabilityai/stable-diffusion-3-medium-diffusers
• LLM: mistralai/Mistral-7B-Instruct-v0.3

License

This project is licensed under the MIT License.

References

If you find our work useful, please consider citing our paper :)

@article{yoon2025splitflow,
  title={SplitFlow: Flow Decomposition for Inversion-Free Text-to-Image Editing},
  author={Yoon, Sung-Hoon and Li, Minghan and Beaudouin, Gaspard and Wen, Congcong and Azhar, Muhammad Rafay and Wang, Mengyu},
  journal={arXiv preprint arXiv:2510.25970},
  year={2025}
}

This project is based on FlowEdit (ICCV 2025), Huge thanks to authors!