Inverse-and-Edit: Effective and Fast Image Editing by Cycle Consistency Models

This is the official implementation of "Inverse-and-Edit: Effective and Fast Image Editing by Cycle Consistency Models"

Recent advances in image editing with diffusion models have achieved impressive results, offering fine-grained control over the generation process. However, these methods are often computationally intensive due to their iterative nature. While distilled diffusion models enable faster inference, their editing capabilities remain limited --- primarily due to poor inversion quality. High-fidelity inversion and reconstruction are essential for precise image editing, as they preserve the structural and semantic integrity of the source image. In this work, we propose a novel framework that enhances image inversion using consistency models, enabling high-quality editing in just four steps. Our method introduces a cycle-consistency optimization strategy that significantly improves reconstruction accuracy and enables a controllable trade-off between editability and content preservation. We achieve state-of-the-art performance across various image editing tasks and datasets, demonstrating that our method matches or surpasses full-step diffusion models while being substantially more efficient.

Installation

git clone https://github.com/ControlGenAI/Inverse-and-Edit.git
conda create -n inverse-and-edit python=3.10
conda activate inverse-and-edit
pip install -r requirements.txt

Downloads

Our method relies on Guidance-distilled Stable Diffusion 1.5 checkpoint and backward iCD checkpoint. We also provide ours forward CM checkpoint.

sh ./checkpoints/download_checkpoints.sh

MS-COCO data for fine-tune and statistic files.

sh training/data/download_coco_train2014.sh
sh training/data/download_coco_val2014.sh
sh training/stats/download_fid_files.sh

Cycle-consistency fine-tune:

We present a novel fine-tune method based on complete image reconstruction process for inversion quality enhancement.

sh training/sh_scripts/run_sd15_lora.sh

Inversion & Editing

Here, you can find examples of inversion and editing by our method.

Image path and source prompt is required for reconstruction task.

python3 inversion.py --image_path './example_images/coffee.jpg' \
                     --src_prompt 'a cup of coffee with drawing of tulip putted on the wooden table' \
                     --output_path 'reconstructed_image.jpg'

For image editing in our method we adopt guidance strategy.

Image path, source prompt, target prompt and config path is required for image editing. Guiders, cfg schedule and noise-rescale setup should be specified in config.yaml.

python3 editing.py --image_path './example_images/coffee.jpg' \
                   --src_prompt 'a cup of coffee with drawing of tulip putted on the wooden table' \
                   --trg_prompt 'a cup of coffee with drawing of lion putted on the wooden table' \
                   --config_path './configs/consistency.yaml' \
                   --output_path 'edited_image.jpg'

We also provide example_notebook.ipynb with examples of image reconstruction, editing by prompt switching and editing with guidance.

Acknowledgments

This repository primarily based on

• Invertible Consistency Distillation
• Guide-and-Rescale

Citation

@misc{beletskii2025inverseandediteffectivefastimage,
      title={Inverse-and-Edit: Effective and Fast Image Editing by Cycle Consistency Models}, 
      author={Ilia Beletskii and Andrey Kuznetsov and Aibek Alanov},
      year={2025},
      eprint={2506.19103},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.19103}, 
}