Introduction

Love the idea of an AI design studio but hate credits, logins and cloud queues? ComfyArt just landed: a free, MIT-licensed AI design canvas that installs as a ComfyUI custom node and runs every image and video model on your own GPU. Developer jtydhr88 (the creator of the canvas app ComfyTV) announced it on October 16, 2026, openly modeled on the commercial design agent Lovart: an infinite canvas, a design agent that works directly on it, and no account, no credits, no cloud.

Primary sources: the ComfyArt GitHub repo and the ComfyArt docs.

What you get

  • Infinite canvas built on tldraw + React, with pages, frames, layers, text, shapes, comments, snapshot history and export
  • Built for posters, social graphics, slide decks and short videos
  • Image Generation and Video Generation modes that work straight away, no agent needed
  • Design agent (off by default) that can look at the canvas, plan a layout, generate and edit images and videos, place text, frames, web pages and slides, ask clarifying questions and remember your preferences
  • MCP server at /comfyart/mcp, so the built-in agent or any local MCP client can see and edit the canvas
  • Assets: brand kits, characters, products and a reference board you can @-mention in chat
  • License: MIT for ComfyArt itself; the bundled tldraw canvas keeps its own license and shows a small development badge on localhost

The models under the hood

Every canvas action runs one of 15 ComfyUI API workflows shipped in the repo's workflows/ folder:

  • Image generation and edits: Qwen Image 2.1 by default (generate, quick edit, edit text, expand/outpaint), plus Z-Image Turbo, FLUX.2 Klein 9B, Qwen Image 2512 and Ideogram 4 as alternate generators
  • Video: MiniMax H3 for animate-image, text-to-video, first/last-frame video and extend
  • Image tools: 4x-UltraSharp upscale, BiRefNet background removal, LaMa eraser, Qwen Image Layered for splitting layers, Qwen Image Edit 2511 with a multiple-angles LoRA
  • Reference modes: line art, depth, pose and more via comfyui_controlnet_aux preprocessors
  • Voice input and text recognition: a local Gemma 4 model

Qwen Image 2.1 alone is enough for generation, quick edit, edit text and expand, so you can start small and add model files feature by feature.

Video tricks worth knowing

  • Ratios 16:9, 9:16, 1:1, 4:3, 3:4 and 21:9, with 5- or 10-second clips and camera motion presets
  • Animate Image loops by default by reusing the first frame as the last frame
  • Storyboard splits one idea into up to 6 scenes (30 seconds total) with consistent characters and style
  • A built-in video editor for split, trim, transitions, subtitles, volume and speed, plus upscale to 1080p, 2K or 4K with audio and frame rate kept

Bring your own agent

ComfyArt doesn't ship its own LLM or store credentials. It drives command-line agents already signed in on your machine:

  • Claude Code (claude)
  • Codex (codex)
  • Qwen Code (qwen)
  • DeepSeek Harness (dsh)
  • Any local OpenAI-compatible endpoint with tool calling

Turn on Enable MCP and Enable Agent in Agent Settings. Keep the project's ComfyArt tab open while the agent works, because the canvas tools run in that tab.

Install it tonight

You need a working ComfyUI with a GPU that can handle the image model you choose. The built frontend ships in dist/, so Node.js is optional.

cd ComfyUI/custom_nodes
git clone https://github.com/jtydhr88/ComfyArt
pip install -r ComfyArt/requirements.txt

Restart ComfyUI, then open http://127.0.0.1:8188/comfyart/ in your browser. Projects save under your ComfyUI user directory in comfyart/projects/.

One gotcha: if you launch ComfyUI with --fast fp8_matrix_mult, unscaled fp8 checkpoints like Qwen Image 2512 turn to noise. ComfyArt includes a ComfyArt Precise UNET Loader node that skips that path for the model it loads.

Conclusion

ComfyArt is exactly the kind of drop the local crowd has been waiting for: a polished, Lovart-style design canvas on top of the ComfyUI you already run, with a real tool-using agent and your choice of Qwen Image, Klein or MiniMax H3 doing the pixels. It's brand new (first commit October 15) and moving fast, so grab the repo, start with Qwen Image 2.1, and add video once your first poster lands.

No local GPU? You can still try FLUX.2 Klein and MiniMax H3 in the browser.

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=comfyart-open-source-local-lovart-comfyui

—Aurelia ♡