ComfyArt: a free, local "Lovart" for ComfyUI. Canvas, design agent, image/video/audio/3D, all on your GPU (MIT)
I've been building ComfyArt: an AI design canvas in the spirit of Lovart, but open source and running fully on your own ComfyUI. No account, no credits, no cloud unless you want it.
jtydhr88/ComfyArt
What it is
- An infinite canvas (tldraw + React) that lives at /comfyart/ inside ComfyUI, with a ComfyArt button in the
- top bar (works in Comfy Desktop too)
- A design agent that actually works on the canvas through MCP tools: it looks at the board, generates, edits, places text and frames, builds slides and web pages
- Every generation runs as a normal ComfyUI workflow on your machine
What you can do on the canvas
- Generate: Qwen Image 2.1, Z-Image Turbo, FLUX.2 Klein, FLUX 3 Klein, Ideogram 4 and more; references,
- structure modes (depth/pose/line art), palettes
- Video: MiniMax H3, from text, first/last frame, up to 9 reference images, or a storyboard; motion control, lip sync
- Image tools: quick edit, upscale, remove background, eraser, inpaint, expand, move object, multi-angle
- (QMA LoRA), split layers, edit text, vectorize
- Edit Elements: turn a flat poster back into an editable artboard (OCR text layers, cut-out objects, clean background)
- Mockups: print a design onto a product photo; it follows folds and shading
- Music, voiceover (voice cloning), SFX, image-to-3D as cards on the same board
- Rich text, pen tool, booleans, auto layout, layers, brand kit; export PNG/JPG/PDF and editable PPTX
Agent: bring your own brain
Claude Code, Codex, Qwen Code, DeepSeek Harness, or any OpenAI-compatible local LLM with tool calling.
ComfyArt never stores credentials. You can also plug the ComfyArt MCP endpoint into Cursor / Claude Desktop /
your own scripts.
No big GPU?
- Spread jobs across other ComfyUI machines on your LAN (Auto picks the least busy one that has the models)
- Or set your own COMFY_API_KEY and run the same workflows on Comfy Cloud
Try it
1. git clone https://github.com/jtydhr88/ComfyArt into ComfyUI/custom_nodes/
2. pip install -r requirements.txt (no Node build needed, dist/ is committed)
3. Restart ComfyUI, click ComfyArt in the top bar
Docs with screenshots for every feature: comfyart.org Repo (MIT): jtydhr88/ComfyArt
#OpenSourceHiggsfield
9 points · 3 comments
Mentions
@xuezailinxin190 可以做哈 ComfyArt#19
ComfyArt: a free, local "Lovart" for ComfyUI. Canvas, design agent, image/video/audio/3D, all on your GPU (MIT) I've been building ComfyArt: an AI design canvas in the spirit of Lovart, but open source and running fully on your own ComfyUI. No account, no credits, no cloud unless you want it. jtydhr88/ComfyArt What it is - An infinite canvas (tldraw + React) that lives at /comfyart/ inside ComfyUI, with a ComfyArt button in the - top bar (works in Comfy Desktop too) - A design agent that actually works on the canvas through MCP tools: it looks at the board, generates, edits, places text and frames, builds slides and web pages - Every generation runs as a normal ComfyUI workflow on your machine What you can do on the canvas - Generate: Qwen Image 2.1, Z-Image Turbo, FLUX.2 Klein, FLUX 3 Klein, Ideogram 4 and more; references, - structure modes (depth/pose/line art), palettes - Video: MiniMax H3, from text, first/last frame, up to 9 reference images, or a storyboard; motion control, lip sync - Image tools: quick edit, upscale, remove background, eraser, inpaint, expand, move object, multi-angle - (QMA LoRA), split layers, edit text, vectorize - Edit Elements: turn a flat poster back into an editable artboard (OCR text layers, cut-out objects, clean background) - Mockups: print a design onto a product photo; it follows folds and shading - Music, voiceover (voice cloning), SFX, image-to-3D as cards on the same board - Rich text, pen tool, booleans, auto layout, layers, brand kit; export PNG/JPG/PDF and editable PPTX Agent: bring your own brain Claude Code, Codex, Qwen Code, DeepSeek Harness, or any OpenAI-compatible local LLM with tool calling. ComfyArt never stores credentials. You can also plug the ComfyArt MCP endpoint into Cursor / Claude Desktop / your own scripts. No big GPU? - Spread jobs across other ComfyUI machines on your LAN (Auto picks the least busy one that has the models) - Or set your own COMFY_API_KEY and run the same workflows on Comfy Cloud Try it 1. git clone https://github.com/jtydhr88/ComfyArt into ComfyUI/custom_nodes/ 2. pip install -r requirements.txt (no Node build needed, dist/ is committed) 3. Restart ComfyUI, click ComfyArt in the top bar Docs with screenshots for every feature: comfyart.org Repo (MIT): jtydhr88/ComfyArt #OpenSourceHiggsfield
ComfyArt, now running on Comfy desktop. jtydhr88/ComfyArt #OpenSourceHiggsfield #ComfyUI t.co
@jtydhr88 仓库 jtydhr88/ComfyArt 文档 comfyart.org