Zechuan Zhang  Ji Xie  Yu Lu  Zongxin Yang  Yi Yang✉ 

ReLER, CCAI, Zhejiang University; Harvard University

^✉Corresponding Author

Arxiv  Huggingface Demo 🤗  Model 🤗  Project Page

Image Editing is worth a single LoRA! We present In-Context Edit, a novel approach that achieves state-of-the-art instruction-based editing using just 0.5% of the training data and 1% of the parameters required by prior SOTA methods. The first row illustrates a series of multi-turn edits, executed with high precision, while the second and third rows highlight diverse, visually impressive single-turn editing results from our method.

🎆 News

• [2025/9/19] 🔥 We have open-sourced our MoE version ICEdit and ckpt. Have a try!🚀
• [2025/9/18] 🌟 ICEdit has been accepted by NeurIPS 2025!🎉 See you in San Diego!
• [2025/8/21] 🌟 We have released an Ascend (Huawei NPU)-powered version of ICEdit. Now you can run ICEdit on Ascend NPU! Many thanks to Zhiyuan！
• [2025/5/16] 🌟 Many thanks to gluttony-10 (十字鱼) for adapting Gradio demo with GGUF quantization, further reducing memory usage to 10GB.
• [2025/5/14] 🔥 With the help of the official comfy-org, we have integrated our ComfyUI nodes into Comfy Registry!
• [2025/5/13] 🔥 We have released the training code! Train your own editing LoRAs now!
• [2025/5/11] 🌟 Great thanks to gluttony-10 (十字鱼) for making a windows gradio demo to use our project on Windows!
• [2025/5/8] 🔥 We have released our official ComfyUI workflow! 🚀 Check the repository and have a try!

Click to expand/collapse news

• [2025/5/8] 🔥 We have added LoRA scale slider in the gradio demo. You can try to discover more interesting demo with different scale!

• [2025/5/7] 🌟 We update some notes when using the ComfyUI workflow to avoid unsatisfactory results!
• [2025/5/6] 🔥 ICEdit currently ranks 2nd on the overall/weekly trending list of Hugging Face space. Thank you all for your support and love!🤗
• [2025/5/5] 🌟 Heartfelt thanks to Datou for creating a fantastic ComfyUI workflow on OpenArt! 🚀 Have a try!
• [2025/5/2] 🌟 Heartfelt thanks to judian17 for crafting an amazing ComfyUI-nunchaku demo! Only 4GB VRAM GPU is enough to run with ComfyUI-nunchaku!🚀 Dive in and give it a spin!
• [2025/4/30] 🔥 We release the Huggingface Demo 🤗! Have a try!
• [2025/4/30] 🔥 We release the paper on arXiv!
• [2025/4/29] We release the project page and demo video! Codes will be made available in next week~ Happy Labor Day!

🎈 Tutorial on Bilibili or Youtube

👑 Feel free to share your results in this Gallery!

• [2025/5/15] 🌟 We find that 啦啦啦的小黄瓜 has made a detailed bilibili tutorial introducing our model! What a great video!
• [2025/5/14] 🌟 We find that Nenly同学 has made a fantastic bilibili tutorial on how to use our repository! Great thanks to him!
• [2025/5/10] 🌟 Great thanks to 月下Hugo for making a Chinese tutorial on how to use our official workflow!
• [2025/5/7] 🌟 Heartfelt thanks to T8star for making a tutorial and ComfyUI workflow on how to increase the editing success to 100%!🚀 Have a try!
• [2025/5/3] 🌟 Heartfelt thanks to softicelee2 for making a Youtube video on how to use our model!

📖 Table of Contents

• 🎆 News
• 🎈 Tutorial on Bilibili or Youtube
  • 👑 Feel free to share your results in this Gallery!
• 📖 Table of Contents
  • 📢 Attention All: Incorrect ComfyUI Workflow Usage Alert!
• 💼 Installation
  • Conda environment setup
  • Download pretrained weights
  • Inference in bash (w/o VLM Inference-time Scaling)
    • For the usage of MoE-LoRA version
  • Inference in Gradio Demo
  • 💼 Windows one-click package
• 🔧 Training
• 🎨ComfyUI Workflow
  • Official ComfyUI-workflow
  • ComfyUI-workflow for increased editing success rate
  • ComfyUI-nunchaku
  • ComfyUI-workflow
• ⚠️ Tips
  • If you encounter such a failure case, please try again with a different seed!
  • ⚠️ Clarification
• 💪 To Do List
• 💪 Comparison with Commercial Models
• 🌟 Star History
• Bibtex

📢 Attention All: Incorrect ComfyUI Workflow Usage Alert!

• We have released our official ComfyUI workflow for proper usage! Check our repository and have a try!

• You need to add the fixed pre-prompt "A diptych with two side-by-side images of the same scene. On the right, the scene is exactly the same as on the left but {instruction}" before inputing the edit instructions, otherwise you may get bad results! (This is mentioned in the paper!, The code for the Hugging Face gradio demo already embeds this prompt. So, you can simply input the editing instructions without additional setup.)
• The width of the input image must resize to 512 (no restriction to height).
• Please use the Normal LoRA not the MoE-LoRA, because the MoE-LoRA cannot be correctly loaded with ComfyUI lora loader.
• 🔥💐🎆 Welcome to share your creative workflows (such as combining Redux, ACE, etc.) in the Issues section and showcase the results! We will include references so that more people can see your creativity.

💼 Installation

Conda environment setup

conda create -n icedit python=3.10
conda activate icedit
pip install -r requirements.txt
pip install -U huggingface_hub

Download pretrained weights

If you can connect to Huggingface, you don't need to download the weights. Otherwise, you need to download the weights to local.

• Flux.1-fill-dev.
• ICEdit-normal-LoRA.
• ICEdit-MoE-LoRA

Note: Due to some cooperation permission issues, we have to withdraw the weights and codes of moe-lora temporarily. What is released currently is just the ordinary lora, but it still has powerful performance. If you urgently need the moe lora weights of the original text, please email the author.

Inference in bash (w/o VLM Inference-time Scaling)

Now you can have a try!

Our model can only edit images with a width of 512 pixels (there is no restriction on the height). If you pass in an image with a width other than 512 pixels, the model will automatically resize it to 512 pixels.

If you found the model failed to generate the expected results, please try to change the --seed parameter. Inference-time Scaling with VLM can help much to improve the results.

python scripts/inference.py --image assets/girl.png \
                            --instruction "Make her hair dark green and her clothes checked." \
                            --seed 304897401 \

Editing a 512×768 image requires 35 GB of GPU memory. If you need to run on a system with 24 GB of GPU memory (for example, an NVIDIA RTX3090), you can add the --enable-model-cpu-offload parameter.

python scripts/inference.py --image assets/girl.png \
                            --instruction "Make her hair dark green and her clothes checked." \
                            --enable-model-cpu-offload

If you have downloaded the pretrained weights locally, please pass the parameters during inference, as in:

python scripts/inference.py --image assets/girl.png \
                            --instruction "Make her hair dark green and her clothes checked." \
                            --flux-path /path/to/flux.1-fill-dev \
                            --lora-path /path/to/ICEdit-normal-LoRA

For the usage of MoE-LoRA version

python scripts/inference_moe.py --image assets/girl.png \
                            --instruction "Make her hair dark green and her clothes checked." \
                            --seed 42 \

python scripts/inference_moe.py --image assets/girl.png \
                            --instruction "Make her hair dark green and her clothes checked." \
                            --enable-model-cpu-offload

python scripts/inference_moe.py --image assets/girl.png \
                            --instruction "Make her hair dark green and her clothes checked." \
                            --flux-path /path/to/flux.1-fill-dev \
                            --lora-path /path/to/ICEdit-MoE-LoRA

Inference in Gradio Demo

We provide a gradio demo for you to edit images in a more user-friendly way. You can run the following command to start the demo.

python scripts/gradio_demo.py --port 7860



## for MoE version
python scripts/gradio_demo_moe.py --port 7860

Like the inference script, if you want to run the demo on a system with 24 GB of GPU memory, you can add the --enable-model-cpu-offload parameter. And if you have downloaded the pretrained weights locally, please pass the parameters during inference, as in:

python scripts/gradio_demo.py --port 7860 \
                              --flux-path /path/to/flux.1-fill-dev (optional) \
                              --lora-path /path/to/ICEdit-normal-LoRA (optional) \
                              --enable-model-cpu-offload (optional) \

## for MoE version
python scripts/gradio_demo_moe.py --port 7860 \
                              --flux-path /path/to/flux.1-fill-dev (optional) \
                              --lora-path /path/to/ICEdit-normal-LoRA (optional) \
                              --enable-model-cpu-offload (optional) \

Or if you want to run the demo on a system with 10 GB of GPU memory, you can download the gguf models from FLUX.1-Fill-dev-gguf, t5-v1_1-xxl-encoder-gguf and pass the parameters during inference, as in:

python scripts/gradio_demo.py --port 7861 \
                              --flux-path models/flux.1-fill-dev \
                              --lora-path models/ICEdit-normal-LoRA \
                              --transformer models/flux1-fill-dev-Q4_0.gguf \
                              --text_encoder_2 models/t5-v1_1-xxl-encoder-Q8_0.gguf \
                              --enable-model-cpu-offload \

Then you can open the link in your browser to edit images.