ComfyArt: Free Local AI Design Canvas for ComfyUI (Install, Models and Agent Guide)

Canva, Lovart and Higgsfield all want a monthly subscription and a credit balance. ComfyArt does the same kind of job for free and runs on your own PC. It is an open-source AI design canvas, released by Terry Jia (jtydhr88, also the author of ComfyTV) under the MIT licence, that installs as a ComfyUI custom node. You get an infinite canvas with artboards, layers and text, plus image, video, music, voice and 3D generation that all run through your local ComfyUI. A built-in design agent can drive the whole canvas through Claude Code, Codex or a local LLM. This guide covers what ComfyArt does, how to install it, which models each tool needs and exactly which folder every file goes in.

What Is ComfyArt?

ComfyArt is a web app that ComfyUI serves at /comfyart/. It looks and feels like a design tool rather than a node graph. You work on pages and artboards, place text and shapes, drop in photos, and call AI tools from menus and cards. Behind every action, ComfyArt sends a ready-made API workflow from its workflows/ folder to ComfyUI’s /prompt endpoint. So you never wire nodes yourself, but everything still runs on your own GPU with the models already in your ComfyUI folder.

Version 1.0 arrived on October 23, 2026, and version 1.1.0 followed on October 27 with more tools and LAN support. The project page with screenshots of every feature is comfyart.org, and the code is on GitHub.

The licence is MIT for ComfyArt itself. The canvas is built on the tldraw SDK, which has its own licence. Served from localhost it runs in development mode and shows a small “Get a license for production” badge. That badge is harmless for personal use on your own PC.

Main Features

ComfyArt packs a lot into one canvas. Here are the parts you will use most.

• Canvas and layout: pages, artboards with size presets for social posts and print, fill, stroke, corner radius, auto layout, a layers panel, comments, snapshot history and export to PNG, JPG, PDF and editable PPTX.
• Text and vectors: rich text with mixed fonts and colours in one box, a font picker that includes your local fonts, gradient and image fills, a pen tool, and boolean union, subtract, intersect and exclude.
• Generate: image generator cards sized in real pixels, with five local image models, references, structure modes (line art, depth, pose), colour palettes and base images.
• Image tools: quick edit, upscale, remove background, eraser, crop with platform presets, inpaint, move object, multi-angle, expand, split layers, edit text, style transfer and vectorize.
• Edit Elements: turns a flat poster back into an editable artboard. It reads the text with OCR, cuts out the product and logo, cleans the background, and lets you edit each part before re-rendering.
• Mockups: prints your design onto a product photo, following folds and shading, then bakes or blends it.
• Audio and 3D: music, voiceovers with voice cloning, sound effects, and image to textured 3D model.
• Assets: a brand kit editor, characters, products and a reference board the agent can reuse.

If you have tried our Freefield guide, think of ComfyArt as the design-studio version of that idea: less of an app store, more of a Canva-style editor with AI built in.

How to Install ComfyArt

You need a working ComfyUI install (portable, manual or Comfy Desktop). No build step is needed because the built web page in dist/ is committed to the repo.

Step 1: Clone into custom_nodes

cd ComfyUI/custom_nodes
git clone https://github.com/jtydhr88/ComfyArt

The pack now sits at ComfyUI/custom_nodes/ComfyArt/.

Step 2: Install the Python requirements

Run this inside the same Python environment that ComfyUI uses:

cd ComfyUI/custom_nodes/ComfyArt
pip install -r requirements.txt

This installs pypdfium2 (for PDF uploads) and rapidocr (for Edit Elements). On the Windows portable build, use the embedded Python instead:

cd ComfyUI_windows_portable
python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyArt\requirements.txt

Step 3: Restart and open the canvas

Restart ComfyUI. Click the new ComfyArt button in ComfyUI’s top bar, or open http://127.0.0.1:8188/comfyart/ in your browser. Back to ComfyUI is in the home rail and in each project’s logo menu.

Your projects, snapshot history and agent chats are stored in a SQLite database at ComfyUI/user/__comfyart/data.db. Back up that one file and you have backed up all your designs.

Updating

cd ComfyUI/custom_nodes/ComfyArt
git pull
pip install -r requirements.txt

Models You Need, With Folder Paths

ComfyArt does not download the big models for you. Each tool needs specific files, and a tool simply fails if its files are missing. You do not need all of them. Start with the default image model, then add the rest as you need each tool.

Default image model: Qwen Image 2.1

Image generation, quick edit, edit text, expand and comment edits all use Qwen Image 2.1:

• ComfyUI/models/diffusion_models/qwen_image_2.1_int8_convrot.safetensors
• ComfyUI/models/text_encoders/qwen3vl_8b_int8_convrot.safetensors
• ComfyUI/models/vae/qwen_image_2.1_vae_bf16.safetensors

Our Qwen Image 2.1 guide for 16 GB cards has the download links and settings for these files.

Other image models

All files end in .safetensors. Each one goes in ComfyUI/models/<column folder>/, for example ComfyUI/models/text_encoders/qwen_3_4b.safetensors. For setup details on two of these, see our FLUX 3 Klein guide and MOSS-Image guide.

Qwen Image 2512 loads with the ComfyArt Precise UNET Loader node that ships with the pack. ComfyUI’s --fast fp8_matrix_mult option turns unscaled fp8 checkpoints like this one into noise, and this loader avoids that path.

Image tool models

The ControlNet-aux annotators for reference modes (MLSD, HED, normal, segmentation) and Depth Anything V2 for mockups download on first use, so there is nothing to place by hand.

Extra Custom Nodes for Some Tools

Most workflows use core ComfyUI nodes. These four tools need an extra pack, each installed into ComfyUI/custom_nodes/:

cd ComfyUI/custom_nodes
git clone https://github.com/Fannovel16/comfyui_controlnet_aux
git clone https://github.com/Acly/comfyui-inpaint-nodes
git clone https://github.com/stavsap/comfyui-kokoro
git clone https://github.com/yolain/ComfyUI-Easy-IndexTTS2

Then install each pack’s requirements and restart:

pip install -r comfyui_controlnet_aux/requirements.txt
pip install -r comfyui-kokoro/requirements.txt
pip install -r ComfyUI-Easy-IndexTTS2/requirements.txt

• comfyui_controlnet_aux: reference modes (line art, depth, pose) and mockups.
• comfyui-inpaint-nodes: the LaMa eraser and Edit Elements background clean-up.
• comfyui-kokoro: voiceovers.
• ComfyUI-Easy-IndexTTS2: voice cloning.

You can also install all four from ComfyUI Manager by searching for their names.

Your First Design, Step by Step

1. Start a project. On the home screen, type one sentence such as “Instagram post for a coffee shop autumn menu”, or pick a skill. ComfyArt creates a project with a sized artboard.
2. Add a generator card. Choose an image generator, set the pixel size to match the artboard, pick Qwen Image 2.1 (or another model you installed) and write the prompt.
3. Add a reference if you want. Drop a photo onto the card and choose a structure mode such as depth or line art to keep the layout.
4. Generate and place. The result lands on the canvas as a layer. Move it into the artboard.
5. Add text. Use the text tool with your brand font. Text stays real text, so it remains sharp and editable, unlike text baked into an AI image.
6. Clean up. Use remove background, eraser or quick edit on any layer. Each tool runs a small workflow and replaces the layer.
7. Export. Export as PNG or JPG for social media, PDF for print, or PPTX if someone wants to edit it in PowerPoint.

A handy trick is the comment tool. Draw a comment on part of the design, write “make this cup red”, and run it as an edit. ComfyArt sends the marked area to the edit model, so you change one spot without regenerating the whole image.

Video, Music, Voice and 3D

Video with MiniMax H3

Animate, video generation and video extend use MiniMax H3. Video lengths are 5 to 15 seconds at 768p or 1080p, from text, first and last frames, up to 9 reference images, or a storyboard. The files are:

• ComfyUI/models/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors
• ComfyUI/models/loras/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors
• ComfyUI/models/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
• ComfyUI/models/vae/minimax_h3_video_vae_fp16.safetensors
• ComfyUI/models/vae/minimax_h3_audio_vae_fp32.safetensors

Reference to video adds ComfyUI/models/diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors. Motion control adds three more: ComfyUI/models/model_patches/minimax_h3_fun_controlnet_union_pruned_int8_convrot.safetensors, ComfyUI/models/checkpoints/sdpose_wholebody_fp16.safetensors and ComfyUI/models/diffusion_models/rt_detr_v4-x-hgnet_fp16.safetensors. Lip sync uses the image-to-video files with the MiniMaxH3AddGuide node.

If MiniMax H3 is new to you, start with our MiniMax H3 image-to-video workflow and the out-of-memory fix for 16 GB cards.

Music

Three music engines are supported. Pick one:

• MiniMax Music 3: ComfyUI/models/diffusion_models/minimax_music3_dit_fp16.safetensors, ComfyUI/models/text_encoders/minimax_music3_text_encoder_pruned_int8_convrot.safetensors, ComfyUI/models/vae/minimax_music3_dav.safetensors
• ACE-Step 1.5: ComfyUI/models/diffusion_models/acestep_v1.5_xl_turbo_bf16.safetensors, ComfyUI/models/text_encoders/qwen_0.6b_ace15.safetensors and qwen_4b_ace15.safetensors, ComfyUI/models/vae/ace_1.5_vae.safetensors
• YuE2: ComfyUI/models/checkpoints/yue2_3b_int8_convrot.safetensors

Sound effects and voice

Sound effects use ComfyUI/models/checkpoints/stable-audio-open-1.0.safetensors and ComfyUI/models/text_encoders/t5-base.safetensors. Voiceovers use the comfyui-kokoro pack, and voice cloning uses ComfyUI-Easy-IndexTTS2, both installed above.

Image to 3D

TRELLIS.2 needs ComfyUI/models/diffusion_models/trellis_2_int8_convrot.safetensors, ComfyUI/models/vae/trellis_2_shape_vae_bf16.safetensors, ComfyUI/models/vae/trellis_2_texture_vae_bf16.safetensors, ComfyUI/models/clip_vision/dino_v3_L_naf_fp32.safetensors and the BiRefNet file. Pixal3D adds ComfyUI/models/diffusion_models/pixal3d_int8_convrot.safetensors and ComfyUI/models/geometry_estimation/moge_2_vitl_normal_fp16.safetensors. Our Pixal3D and TRELLIS.2 guide covers the 3D workflow in more detail.

The Design Agent and MCP

The agent panel is ComfyArt’s headline feature. It runs a real tool-using agent that can look at the canvas, generate and edit images, video, audio and 3D, run Edit Elements and mockups, place text, shapes and frames, and ask you questions. It is off by default. Turn on MCP and the agent in Agent Settings. Image and video tools work without it.

You pick the agent per chat:

For a fully local setup, point the Local LLM URL at Ollama (http://127.0.0.1:11434/v1) or a llama.cpp server running a model that handles tool calls well. If the endpoint needs a key, set COMFYART_LOCAL_LLM_API_KEY before you start ComfyUI. ComfyArt does not store credentials.

Keep the ComfyArt tab of the project open while the agent works, because canvas tools run inside that tab.

Use ComfyArt from your own agent

With MCP turned on, any MCP client can drive a project. Replace <project-id> with your project’s id:

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>"

The MCP guide on comfyart.org has setup for Gemini CLI, Cursor, VS Code and Claude Desktop too. This works much like the official Comfy Agent, but on a design canvas instead of a node graph.

Comfy Cloud and LAN Machines

No GPU: run on Comfy Cloud

Create an API key at platform.comfy.org (Cloud API keys need a paid Comfy plan). Set it as an environment variable before starting ComfyUI:

# Linux / macOS
export COMFY_API_KEY=your_key_here
python main.py

# Windows (Command Prompt)
set COMFY_API_KEY=your_key_here
python main.py

Then choose Settings, Generate on, Comfy Cloud. Cloud has the models and nodes preinstalled, so you download nothing. On Cloud, voiceover and cloning use Qwen3-TTS instead of Kokoro and IndexTTS2. To compare costs with a local card, see our cloud GPU pricing vs RTX 5080 breakdown.

Several PCs: LAN instances

Other ComfyUI machines on your network can share the work. Start each with python main.py --listen, then add them under Settings, LAN instances (name, address, port). “Auto” sends each job to the least busy machine that has every node and model the workflow needs. Model files are matched by file name, so they can sit in different subfolders on each machine. Qwen Image 2512 also needs ComfyArt installed on the remote machine for its loader.

How Much VRAM Do You Need?

ComfyArt adds almost nothing on top of ComfyUI. What matters is which model each tool loads. The defaults are int8 “convrot” files, chosen to keep memory low.

Estimate for a 16 GB card like the RTX 5080: based on the int8 file sizes and ComfyUI’s offloading, every ComfyArt tool should run on 16 GB of VRAM with 32 GB or more of system RAM. Image tools should feel close to interactive, while 1080p video clips take minutes rather than seconds. This is an estimate, not a measured result. Our VRAM guide for ComfyUI explains how offloading trades speed for memory.

ComfyArt vs Freefield vs Canva

If you mostly want quick one-off generations, Freefield is simpler. If you build posters, thumbnails, slides or social posts with real text and layers, ComfyArt is the better fit. For GIMP users, OpenFill brings generative fill to a classic editor instead.

Troubleshooting

• No ComfyArt button in the top bar: check that the folder is ComfyUI/custom_nodes/ComfyArt/ (not nested one level deeper) and look in the ComfyUI console for import errors. You can always open http://127.0.0.1:8188/comfyart/ directly.
• A tool fails at once: the workflow is missing a model or node. The ComfyUI console names the missing file. Put it in the folder listed above and press refresh in ComfyUI.
• Qwen Image 2512 makes pure noise: you are running with --fast fp8_matrix_mult and the model was not loaded by the ComfyArt Precise UNET Loader. Update ComfyArt, or start ComfyUI without that flag.
• Edit Elements or PDF upload errors: rapidocr or pypdfium2 is missing. Re-run pip install -r requirements.txt in ComfyUI’s Python.
• Agent does nothing: MCP and the agent are off by default. Turn both on in Agent Settings and check that claude, codex or your local endpoint works in a terminal first.
• Out of memory on video: lower the resolution to 768p, shorten the clip, and use the turbo LoRA.

FAQ

Is ComfyArt free?

Yes. ComfyArt is MIT licensed and runs on your own hardware. The only paid option is Comfy Cloud if you have no GPU.

Does ComfyArt work with Comfy Desktop?

Yes. The ComfyArt button appears in Comfy Desktop’s top bar too. Install it into the Desktop app’s custom_nodes folder and its Python environment.

Do I need every model in the list?

No. Install the Qwen Image 2.1 files first. Add the other models only for the tools you plan to use.

Can I use it commercially?

ComfyArt’s code is MIT, which allows commercial use. Each AI model has its own licence, so check the licence of the image, video and audio models you use. The tldraw canvas shows a development badge on localhost; a production deployment of tldraw needs its own licence.

Can ComfyArt run without the internet?

Yes, once the models and annotators are downloaded. The only parts that need the internet are Comfy Cloud and cloud agents like Claude Code or Codex. A local LLM keeps the agent offline too.

Where are my projects saved?

In ComfyUI/user/__comfyart/data.db. Generated files go to ComfyUI/output/comfyart/.

Final Thoughts

ComfyArt is the most complete free design canvas built on ComfyUI so far. It turns a node-based tool into something a non-technical designer can use, while keeping every model local. The setup work is all in the model downloads, so start small: install the pack, add the three Qwen Image 2.1 files, make one social post, then add video, music and 3D as you need them. If you already run ComfyUI, it costs you nothing to try.

More ComfyUI guides:

• Freefield: Free Open-Source Higgsfield Alternative
• OpenFill: Free GIMP Generative Fill With ComfyUI
• Best Qwen Image 2.1 Turbo Workflow for 16 GB
• Comfy Agent Guide

Sources

• GitHub: jtydhr88/ComfyArt (README, read November 1, 2026)
• comfyart.org (official documentation)
• comfyart.org: Models guide (model folders)
• comfyart.org: MCP guide
• tldraw SDK