DirectEdit: Step-Level Accurate Inversion for Flow-Based Image Editing

TL;DR: DirectEdit is a simple yet effective training-free editing method that eliminates the inherent reconstruction error of inversion, achieving precise background preservation and reliable feature sharing.

🔧 Setup

Install the conda environment:

conda env create -f environment.yml
conda activate directedit

Download the checkpoints SD-3.5-medium, FLUX.1 Dev and SAM2 on Hugging Face.

🚀 Inference on Real Image

You can edit your own images using the Gradio demo:

python demo.py

or you can edit a single image with the following script:

python edit_real_flux_singleimg.py --recov_cfg 3\
                         --attn_ratio 0.1\
                         --src_path 'examples/1.jpg'\
                         --src_prompt 'a cup of coffee with a drawing of a tulip put on the wooden table.'\
                         --tar_prompt 'a cup of coffee with a drawing of a lion put on the wooden table.'\
                         --saved_path './'\
                         --mask_path 'mask_path'(Optional)

python edit_real_sd35_singleimg.py --recov_cfg 2\
                         --attn_ratio 0.3\
                         --src_path 'examples/1.jpg'\
                         --src_prompt 'a cup of coffee with a drawing of a tulip put on the wooden table.'\
                         --tar_prompt 'a cup of coffee with a drawing of a lion put on the wooden table.'\
                         --saved_path './'\
                         --mask_path 'mask_path'(Optional)

Intuition of hyperparameters:

recov_cfg

• Recommended: 2.0 - 3.5
• The higher the CFG scale, the more closely the generated image aligns with the prompt. Values too small (<1.5) or too large (>4.0) will produce poor quality results.

attn_ratio

• Recommended: 0.05 - 0.3
• This hyperparameter influences the consistency between the edited region and the source image, where a greater number of injection steps results in higher similarity.

mask_path

• Effect: Achieving precise background preservation using masks.
• The mask can be obtained through segmentation or manual drawing. We provide a mask segmentation function in the Gradio demo; you can obtain the mask by clicking the edit object of the uploaded image. Alternatively, you can utilize scripts/generate_mask.py to automatically generate masks based on MLLM and SAM (requires configuring an API key).

We also provide batch scripts based on the PIE-bench for evaluation:

# Set python path
export PYTHONPATH=$(realpath "./"):$PYTHONPATH

# SD3.5 
python scripts/edit_real_sd35.py --inv_cfg 1 --recov_cfg 2 --attn_ratio 0.3 --src_path "PIE-bench" --saved_path "outputs/sd35"

# FLUX.1-dev
python scripts/edit_real_flux.py --inv_cfg 1 --recov_cfg 2 --attn_ratio 0.15 --src_path "PIE-bench" --saved_path "outputs/flux"

# Evaluation
python evaluation/evaluate.py --metrics "structure_distance" "psnr_unedit_part" "lpips_unedit_part" "mse_unedit_part" "ssim_unedit_part" "clip_similarity_source_image"  "clip_similarity_target_image" "clip_similarity_target_image_edit_part"\
 --result_path evaluation.csv\
 --edit_category_list 0 1 2 3 4 5 6 7 8 9 --tar_image_folder "outputs/flux"\
 --tar_method "flux_directedit"\
 --src_image_folder "PIE-bench/annotation_images"\
 --annotation_mapping_file "PIE-bench/mapping_file.json"

📖 Citation

If you find our work helpful, please cite our paper. Thanks for your support!

@article{yang2026directedit,
  title={DirectEdit: Step-Level Accurate Inversion for Flow-Based Image Editing},
  author={Yang, Desong and Ye, Mang},
  journal={arXiv preprint arXiv:2605.02417},
  year={2026}
}

Acknowledgements

The code is largely based on FTEdit. Special thanks to the authors for making their code public!

Contact

If you have any questions or concerns, please contact: desong.yang@whu.edu.cn.