ant-research/magicquill

MagicQuill is an image editing tool that lets you make precise, localized changes to photos and illustrations by drawing directly on them. Instead of describing what you want in text alone, you draw colored strokes on the parts of the image you want to change, and the system figures out what you are trying to do. It was accepted at CVPR 2025, a major computer vision research conference, and is developed by researchers from HKUST, Ant Group, ZJU, and HKU. The core idea is that two types of strokes trigger different effects: adding new content or changing existing content. An AI model watches what you draw and tries to guess what you mean, suggesting text prompts automatically through a feature called DrawNGuess. You can accept, ignore, or modify those suggestions before the edit is applied. This means you spend less time writing detailed descriptions and more time pointing at what you actually want to change. To run the tool on your own computer, you need a graphics card with at least 8 GB of video memory. Setup involves cloning the repository with a specific flag to include a submodule, downloading about 25 GB of model checkpoints, and installing several Python packages including a custom interface component. There are setup scripts for both Windows and Linux to make this easier. Once running, the tool opens in a browser through Gradio, a Python library for building simple web interfaces. For people without a powerful GPU, the project also provides a hosted demo on Hugging Face Spaces and Modelscope so you can try it without any local installation. A ComfyUI extension is available for users who prefer that workflow. The project has been updated steadily since late 2024, adding features like drag-and-drop uploads, auto-save, resolution adjustment, and Docker container support. A second version, MagicQuillV2, has since been released and is linked from the README for anyone wanting the latest iteration.