What it does MagicQuill is a local image editing system with a Gradio interface and three brush modes: add elements, subtract/remove them, and color specific regions. An MLLM backend attempts to auto-complete your intent from the brush strokes, so you don’t have to craft elaborate prompts. It also comes as a ComfyUI custom node and a Docker container.

The interesting bit The “DrawNGuess” feature is the quiet workhorse: it watches where you paint and suggests what you’re trying to do, filling in the prompt gap that usually stalls diffusion-based editors. The README is unusually honest about hardware limits—8GB VRAM minimum, with a warning to disable auto-prompting if your GPU chokes.

Key highlights

• Three semantic brushes (add / subtract / color) with edge-aware masking
• Automatic prompt generation via multimodal LLM; wand icon toggles it off
• Runs locally on ~8GB VRAM (tested on RTX 4070 Laptop), with Hugging Face and ModelScope fallbacks
• 25GB of checkpoints required; Docker and Windows batch scripts provided
• ComfyUI node available separately; V2 already released at separate repo

Caveats

• Setup is involved: recursive git clone, manual checkpoint download, LLaVA submodule wrangling, and a pinned PyTorch 2.1.2/CUDA 11.8 stack
• The 25GB model download is hosted on SharePoint with a fallback to HuggingFace; the primary link may be flaky
• V2 exists at a different repository, suggesting this version is already in maintenance mode

Verdict Worth trying if you do a lot of local inpainting and hate prompt engineering. Skip if you’re hoping for a lightweight drop-in; the setup friction and checkpoint bulk are real.