ComfyArt is an open-source design canvas that runs on your own ComfyUI
ComfyArt is an MIT-licensed design canvas that installs into ComfyUI and lets a design agent create posters, slides, mockups, videos, music and 3D models locally, with 82 GitHub stars in its first five days.
Image: GitHub
Why it mattersA designer or marketer who was paying a Lovart subscription for a cloud design canvas can run the same shape of workflow on their own machine, with no account and no per-image credit cost.
A designer who wants an AI design canvas has had to pay a cloud service for every image and video the canvas generates. ComfyArt, an MIT-licensed project by jtydhr88, installs into ComfyUI and runs the whole canvas locally, with no account, no credits and no cloud call. The repository is five days old and already carries 82 GitHub stars.
The maintainer's README calls ComfyArt a free, open-source Lovart alternative, pointing at the paid design agent whose canvas-plus-chat shape it reproduces. The canvas itself is tldraw plus React, so it has pages, artboards with size presets, a layers panel, snapshot history and export to PNG, JPG, PDF and editable PPTX. A chat side panel holds the design agent, which can read the canvas and act on it through an MCP server the add-on exposes.
What the canvas can do
The feature list in the README covers more than image generation. Images come from five local models that the user picks for the pixel size they want: Qwen Image 2.1, Z-Image Turbo, FLUX 2 Klein, Qwen Image 2512 and Ideogram 4. Video is text-to-video, first-and-last-frame or a storyboard, five to fifteen seconds at 768p or 1080p, using MiniMax H3 locally. Audio has music through ACE-Step, sound effects through Stable Audio, voiceover through Kokoro and voice cloning through IndexTTS2. Image-to-3D goes through Trellis 2.
Beyond generation, the canvas has image tools that read as a Photoshop subset: upscale, remove background, inpaint, crop with platform presets, expand, split layers, style transfer and vectorize. An Edit Elements tool reads a flat poster back into an editable artboard by running OCR on the text, cutting out products and logos and cleaning the background, so a design that arrived as a JPG can be edited and re-rendered.
The agent and which models can drive it
The agent side runs any of Claude Code, Codex, Qwen Code, DeepSeek Harness or an OpenAI-compatible local model. The maintainer requires each one on PATH and logged in, and ComfyArt does not store credentials. The agent is off by default, so image and video modes work on their own; turning on MCP in Agent Settings gives the model access to every tool on the canvas, including placing text, shapes, frames, web pages and slides.
The maintainers do not quote a benchmark, a quality comparison with Lovart or a measured time-to-task. The models the canvas depends on are a long list, each with its own licence, and a user has to download each one into ComfyUI before the matching action works. The tldraw canvas carries its own licence, bundled in the dist directory, and shows a small "Get a license for production" badge when ComfyArt is served from a non-local address.
For a designer who already runs ComfyUI, the trade is one add-on install against the model weights and the time to set up each action's workflow. For a designer who does not run ComfyUI, the barrier is higher: this is a canvas on top of a stack, not a self-contained application. Full documentation with screenshots is at comfyart.org.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.