ComfyArt

English | 中文

Watch the full promo (38s)

A free, open-source AI design canvas that runs fully on your own ComfyUI: an infinite canvas (tldraw + React), a design agent that works on it with tools, and every image, video, audio and 3D model running locally. No account, no credits, no cloud.

Documentation with screenshots for every feature: comfyart.org.

• Home and projects: start a design from one sentence or a skill, browse and manage projects, switch language (中文 / English), theme and chat side at any time.
• Canvas: pages, artboards with size presets, fill, stroke and corner radius, auto layout, a layers panel with drag, rename and range select, comments that can be run as edits or sent to the Agent with the marked area, snapshot history, export to PNG, JPG, PDF and editable PPTX.
• Text and vectors: rich text (mixed fonts, sizes, colors, bold, italic, underline in one text box), a font picker with brand, common and local fonts, solid, linear, radial and image fills with transparent color stops and blend modes, a pen tool with node editing, and boolean union, subtract, intersect and exclude.
• Generate: image generators sized in real pixels with five local image models, references and structure modes, color palettes and base images; video from text, first and last frames, up to 9 reference images, or a storyboard, 5 to 15 seconds at 768p or 1080p.
• Image tools: quick edit, upscale, remove background, eraser, crop with platform presets, flip and rotate, non-destructive adjust, inpaint, move object, multi-angle, expand, split layers, edit text, style transfer, vectorize.
• Edit Elements: turn a flat poster back into an editable artboard (OCR text layers, cut-out product and logo, clean background), edit it, then re-render.
• Mockups: print a design onto a product photo; it follows folds and shading in real time, then bake it or blend it with AI.
• Audio and 3D: music, voiceovers (with voice cloning) and sound effects as playable cards; image to textured 3D model.
• Assets: My Assets, a full-page brand kit editor, characters, products and a reference board.
• Agent: a tool-using design agent (below) that can use every feature above through MCP.

Put this folder in ComfyUI/custom_nodes/, install the Python requirements into ComfyUI's environment (pip install -r requirements.txt: pypdfium2 for PDF uploads and rapidocr for Edit Elements), restart ComfyUI and open http://127.0.0.1:8188/comfyart/. The built page is in dist/; after changing src/, rebuild with npm install && npm run build.

Projects, their snapshot history and agent chats are saved in a SQLite database, user/__comfyart/data.db. Projects from older versions (JSON files under comfyart/projects/) are imported automatically the first time you open ComfyArt; the files are left in place. The interface starts in ComfyUI's language and can be switched from the settings menu.

ComfyArt can run its workflows on Comfy Cloud instead of the local GPU. Create an API key at platform.comfy.org (Cloud API keys need a paid Comfy plan), set it as COMFY_API_KEY before starting ComfyUI, then choose Settings → Generate on → Comfy Cloud. The same workflows are sent to https://cloud.comfy.org/api/v2/jobs; no build or deployment is needed because Cloud has the models and nodes preinstalled. Input images and audio are uploaded as Cloud assets when a job is submitted, and results are downloaded into output/comfyart/ with a cloud-<job> suffix, so the canvas, history and Agent see them like local results.

On Cloud, FLUX.2 Klein 9B runs the bf16 flux-2-klein-9b.safetensors and Qwen Image 2512 loads with the core UNETLoader. Angles imports the QMA LoRA from Hugging Face into your Cloud model library the first time it runs. Voiceover and voice cloning use Qwen3-TTS on Cloud (speech-qwen3tts, voice-clone-qwen3tts) instead of Kokoro and IndexTTS2, which Cloud does not have.

Other ComfyUI machines on your network can run ComfyArt jobs too. Start them with --listen, and add them under Settings → LAN instances… (name, address, port; nothing else is stored). Settings → Generate on then sets where jobs go by default: Auto sends each job to the least busy machine, this computer included, that has every node and model file the workflow needs, skipping machines that are off or missing something; picking a machine queues jobs there. Each generator card can override it under Run on, next to the model. Tools such as upscale or remove background use the default, and the card shows where a running job went. Input files are uploaded to the machine's input/comfyart/, progress shows on the canvas as usual, and results are downloaded into output/comfyart/ with a <machine>-<job> suffix. Model files are matched by file name, so they can sit in different subfolders on each machine. Qwen Image 2512 needs ComfyArt installed on that machine for its loader.

The Agent panel runs a real tool-using agent on the backend, ported from ComfyTV's bot. It is off by default: turn on MCP and the Agent in Agent Settings. Image and video modes work without it. Pick the agent per chat:

Storyboard split, OCR, brand and memory extraction, reference notes and the Ideogram caption are single calls to the default Agent (POST /comfyart/bot/complete, no tools). ComfyArt never stores credentials. The agent works through the ComfyArt MCP server at POST /comfyart/mcp (look at the canvas, generate and edit images, videos, audio and 3D models, Edit Elements and mockups, place text, shapes, frames, web pages and slides, memory, skills, ask the user). Canvas tools are executed by the open ComfyArt tab of the project, so keep it open while the agent works. Chats are saved in a SQLite database at user/__comfyart/data.db and keep running when the panel is closed.

Turn on MCP in Agent Settings and add http://127.0.0.1:8188/comfyart/mcp?project=<project-id> to any MCP client, for example:

claude mcp add --transport http comfyart "http://127.0.0.1:8188/comfyart/mcp?project=<project-id>"
codex mcp add comfyart --url "http://127.0.0.1:8188/comfyart/mcp?project=<project-id>"

Setup for Claude Code, Codex, Gemini CLI, Qwen Code, Cursor, VS Code, Claude Desktop and your own scripts, tool timeouts and the full tool list are in the MCP guide.

Every canvas action runs one of the API workflows in workflows/ through ComfyUI's /prompt. They use core nodes unless noted.

• ComfyArt Precise UNET Loader loads a diffusion model with manual cast ops. ComfyUI's --fast fp8_matrix_mult turns unscaled fp8 checkpoints such as Qwen Image 2512 into noise; this loader skips that path for the model it loads.
• /comfyart/bot/* (chats, send, stop, branch, answers), /comfyart/settings and /comfyart/mcp serve the Agent.
• POST /comfyart/cloud/prompt, GET /comfyart/cloud/history/{job} and POST /comfyart/cloud/cancel/{job} forward jobs to Comfy Cloud with COMFY_API_KEY.
• /comfyart/instances lists, adds, edits and removes LAN instances; /comfyart/instances/status reports which are online with their queue and diffusion models. POST /comfyart/remote/prompt picks a machine (target is an instance id or auto) and queues the job there; GET /comfyart/remote/history/{instance}/{job} and POST /comfyart/remote/cancel/{instance}/{job} follow it.
• POST /comfyart/pdf renders up to 30 pages of an uploaded PDF to PNG files in input/comfyart/pdf/ and returns them with each page's text.
• POST /comfyart/ocr reads the text lines of an image (RapidOCR, on device) and returns each line with its box.

npm run dev starts Vite with /api proxied to http://127.0.0.1:8188 (override with COMFY_URL).

ComfyArt is MIT licensed. The Agent backend in bot/ and api/ is adapted from ComfyTV's bot (MIT), including providers co-authored by Pengjingyu1992 (DeepSeek Harness, Local LLM, Codex), MuAIGC (Codex, Qwen Code) and Goldlionren (Codex). The canvas is the tldraw SDK, which is not: it is bundled in dist/ under its own license, copied in licenses/tldraw-LICENSE.md. Served from localhost or plain http it runs in development mode without a license key and shows a small "Get a license for production" badge.