MagicQuill

Brush strokes drive everything in MagicQuill: an add brush sketches new objects, a subtract brush removes them, and a color brush repaints regions, while a multimodal model based on LLaVA watches the strokes and predicts editing intent, proposing the text prompt automatically so editing feels like drawing rather than prompt engineering. The Ant Research system, accepted to CVPR 2025, executes edits through a BrushNet dual-branch inpainting pipeline on Stable Diffusion 1.5 checkpoints such as Realistic Vision and DreamShaper, orchestrated over ComfyUI internals. Running it locally means a conda, scripted, or Docker setup, roughly 25 GB of checkpoint downloads, and a GPU with at least 8 GB of VRAM, confirmed by the authors on an RTX 4070 laptop card; hosted demos on magicquill.art, Hugging Face Spaces, and ModelScope cover lighter hardware. An official ComfyUI node slots it into larger workflows, and MagicQuillV2 introduced layered visual cue editing in late 2025. The repository is licensed CC BY-NC 4.0, so commercial use requires separate permission from the authors.