📝 Introduction

Inversion-based image editing is rapidly gaining momentum while suffering from significant computation overhead, hindering its application in real-time interactive scenarios. In this paper, we rethink that the redundancy in inversion-based image editing exists in both the spatial and temporal dimensions, such as the unnecessary computation in unedited regions and the redundancy in the inversion progress.

To tackle these challenges, we propose a practical framework, named EEdit, to achieve efficient image editing. Specifically, we introduce three techniques to solve them one by one:

• For spatial redundancy, spatial locality caching is introduced to compute the edited region and its neighboring regions while skipping the unedited regions
• Token indexing preprocessing is designed to further accelerate the caching
• For temporal redundancy, inversion step skipping is proposed to reuse the latent for efficient editing

Our experiments demonstrate an average of 2.46X acceleration without performance drop in a wide range of editing tasks including prompt-guided image editing, dragging and image composition.

🛠️ Installation

conda create -n eedit python=3.12.3
conda activate eedit
pip install -r EEdit/requirements.txt

📥 Checkpoints & Datasets

All model weights and datasets are from open-source, free and publicly available channels:

We use FLUX-dev as our experimental model. You can obtain it from either:

• Official Repository: https://github.com/black-forest-labs/flux
• Hugging Face: https://huggingface.co/black-forest-labs/FLUX.1-dev

We use PIE-BENCH as the prompt-guided dataset, you can refer to link

We use TF-ICON benchmark as the ref-guided dataset, you can refer to link

We use DragBench-DR and Drag-Bench-SR as the drag-guided datasets, you can refer to link

For masks generated from mapping_file.json, we provide scripts that follows:

python MyCodes/myutils.py

When all the data and checkpoints are ready, please follow our file directory structure to be compatible with the scripts.

weights
├── flux1-dev.safetensors
├── model_index.json
├── scheduler
│   └── scheduler_config.json
├── sd_vae_ft_mse
│   ├── config.json
│   └── diffusion_pytorch_model.safetensors
├── text_encoder
│   ├── config.json
│   └── model.safetensors
├── text_encoder_2
│   ├── config.json
│   ├── model-00001-of-00002.safetensors
│   ├── model-00002-of-00002.safetensors
│   ├── model.safetensors.index-1.json
│   └── model.safetensors.index.json
├── tokenizer
│   ├── merges-1.txt
│   ├── merges.txt
│   ├── special_tokens_map-1.json
│   ├── special_tokens_map.json
│   ├── tokenizer_config-1.json
│   ├── tokenizer_config.json
│   ├── vocab-1.json
│   └── vocab.json
├── tokenizer_2
│   ├── special_tokens_map-1.json
│   ├── special_tokens_map.json
│   ├── spiece.model
│   ├── tokenizer.json
│   └── tokenizer_config.json
├── transformer_config.json
└── vae
    ├── config.json
    └── diffusion_pytorch_model.safetensors

input
├── composition
│   ├── Real-Cartoon
│   │   ├── 0000 a cartoon animation of a sheep in the forest
│   │   │   ├── bg58.png
│   │   │   ├── cp_bg_fg.jpg
│   │   │   ├── dccf_image.jpg
│   │   │   ├── fg35_63d22cda1f5b66e8e5aca776.jpg
│   │   │   ├── fg35_mask.png
│   │   │   └── mask_bg_fg.jpg
│   ├── Real-Painting
│   ├── Real-Sketch
│   ├── Real-Real
    ...

├── drag_data
│   ├── dragbench-dr
│   │   ├── animals
│   │   │   ├── JH_2023-09-14-1820-16
│   │   │   │   ├── meta_data.pkl
│   │   │   │   ├── meta_data_region.pkl
│   │   │   │   ├── original_image.png
│   │   │   │   └── user_drag.png
│   └── dragbench-sr
│       ├── art_0
│       │   ├── meta_data.pkl
│       │   ├── meta_data_region.pkl
│       │   ├── original_image.png
│       │   └── user_drag.png
    ...

├── inpaint
│   ├── annotation_images 
│   │   ├── 0_random_140
│   │   │   ├── 000000000000.jpg
│   │   │   
│   │   ├── 1_change_object_80
│   │   │   ├── 1_artificial
│   │   │   │   ├── 1_animal
│   │   │   │   │   ├── 111000000000.jpg
│   │   │   │   ├── 2_human
│   │   │   │   │   ├── 112000000000.jpg
│   │   ├── 2_add_object_80
│   │   │   ├── 1_artificial
│   │   │   │   ├── 1_animal
│   │   │   │   │   ├── 211000000000.jpg
        ...
│   ├── mapping_file.json
│   └── masks
│       ├── mask-000.png
|       ├── mask-001.png
|       ├── ...

🚀 Generation

cd EEdit && source run_gen.sh

✨ Hyper-parameters Guidance

As we all know, edited results are affected by many parameters and even random seed. We try our best to explain those parameters and options that may affect image quality, so that you can generate satisfactory image editing results by yourself.

📝 TODO List

• Release evaluation code
• Release notebook and Hugging Face demo
• Develop more user-friendly interaction logic and experience, including Gradio interface

🙏 Acknowledgements

• Thanks to ToCa for cache implementations
• Thanks to Diffusers for pipeline implementations
• Thanks to Region Drag for dragging implementations

📝 BibTeX

@misc{yan2025eeditrethinkingspatial,
      title={EEdit : Rethinking the Spatial and Temporal Redundancy for Efficient Image Editing}, 
      author={Zexuan Yan and Yue Ma and Chang Zou and Wenteng Chen and Qifeng Chen and Linfeng Zhang},
      year={2025},
      eprint={2503.10270},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.10270}, 
}

📧 Contact

yzx_ustc@mail.ustc.edu.cn