🪶 MagicQuill V2: Precise and Interactive Image Editing with Layered Visual Cues

Paper: MagicQuillV2: Precise and Interactive Image Editing with Layered Visual Cues
Project Page: https://magicquill.art/v2/
Code Repository: https://github.com/zliucz/MagicQuillV2
Hugging Face Spaces Demo: https://huggingface.co/spaces/AI4Editing/MagicQuillV2

TLDR: MagicQuill V2 introduces a layered composition paradigm to generative image editing, disentangling creative intent into controllable visual cues (Content, Spatial, Structural, Color) for precise and intuitive control.

Hardware Requirements

Our model is based on Flux Kontext, which is large and computationally intensive.

VRAM: Approximately 40GB of VRAM is required for inference.
Speed: It takes about 30 seconds to generate a single image.

Important: This is a research project focused on pushing the boundaries of interactive image editing. If you do not have sufficient GPU memory, we recommend checking out our MagicQuill V1 or trying the online demo on Hugging Face Spaces.

Setup

Clone the repository

git clone https://github.com/magic-quill/MagicQuillV2.git
cd MagicQuillV2

Create environment

conda create -n MagicQuillV2 python=3.10 -y
conda activate MagicQuillV2

Install dependencies
```
pip install -r requirements.txt
```
Download models Download the models from Hugging Face and place them in the models/ directory.
```
huggingface-cli download LiuZichen/MagicQuillV2-models --local-dir models
```
Run the demo
```
python app.py
```

System Overview

The MagicQuill V2 interactive system is designed to unify our layered composition framework.

Key Upgrades from V1

Toolbar (A): Features a new Local Edit Brush for defining the target editing area, along with tools for sketching edges and applying color.
Visual Cue Manager (B): Holds all content layer visual cues (foreground props) that users can drag onto the canvas to define what to generate.
Image Segmentation Panel (C): Accessed via the segment icon, this panel allows precise object extraction using SAM (Segment Anything Model) with positive/negative dots or bounding boxes.

Citation

If you find MagicQuill V2 useful for your research, please cite our paper:

@article{liu2025magicquillv2,
  title={MagicQuill V2: Precise and Interactive Image Editing with Layered Visual Cues},
  author={Zichen Liu, Yue Yu, Hao Ouyang, Qiuyu Wang, Shuailei Ma, Ka Leong Cheng, Wen Wang, Qingyan Bai, Yuxuan Zhang, Yanhong Zeng, Yixuan Li, Xing Zhu, Yujun Shen, Qifeng Chen},
  journal={arXiv:2512.03046},
  year={2025}
}

Downloads last month: -; Downloads are not tracked for this model. How to track