modal-comfyui

Run ComfyUI on Modal.com with auto-scaling, GPU snapshots, and easy model management.

Good for testing Wan2.2/LTX-2.5/MiniMax-H3 or other image/video/audio generation models.

Note: If you want to work (creating/editing) on your workflows locally while seamlessly runs the workflow on Cloud GPU, you can install ComfyUI-Proxy custom node on your local ComfyUI https://github.com/anr2me/comfyui-proxy
(ComfyGPU's URL for Remote GPU URL, and optionally ComfyCPU's URL for Remote CPU URL).

Prerequisites

• A Modal account
• Python installed
• uv installed

Installation

1. Clone this repository:

   git clone https://github.com/anr2me/modal-comfyui.git
   cd modal-comfyui

2. Install the Modal client:

   uv sync

3. Set up your modal account (if not done already):

   modal setup

Configuration

Models

Copy models.example.py to models.py and edit it to manage your models. You can specify:

• Hugging Face models(models) using repo_id and filename. Set your HF_TOKEN in huggingface-secret Secrets for a faster download speed and gated models.
• External models(models_ext, e.g. civitai) using a direct url. You can also set your CIVITAI_TOKEN in custom-secret Secrets to download gated models.

Models are downloaded to persistent volumes and symlinked to the specified model_dir.

model_dir accepts two styles:

• Relative path (recommended for standard ComfyUI folders): resolved under /root/comfy/ComfyUI/models/. e.g. "checkpoints" → /root/comfy/ComfyUI/models/checkpoints.
• Absolute path: used as-is. Use this when the target lives outside ComfyUI/models/ (e.g. a custom node's own model directory).

See models.example.py for reference.

Plugins and Custom Nodes

Copy plugins.example.py to plugins.py and edit it to add custom node IDs or titles to be installed via comfy-cli.

• Workflow Dependencies: If you have a workflow_api.json in the root directory, the setup will automatically install the necessary custom nodes for that workflow.

In case of Insufficient Custom Node

Open ComfyUI manager on comfyui and click "Used in Workflow" to see which custom nodes are used in the workflow.
*Note: We are using Legacy ComfyUI Manager.

• Add these custom nodes to comfy_plugins in plugins.py(be careful of node id). You can find the node id at https://registry.comfy.org/
• You can also install custom nodes repository using git url. Add the url, branch, and their dependencies to comfy_plugins_ext in plugins.py (be careful of dependency conflicts).

Usage

Serve (Development)

Run the following command to start ComfyUI in development mode:

modal serve comfyui.py

This will provide a temporary URL where you can access the ComfyUI interface.

Deploy (Production)

To deploy ComfyUI as a persistent app using the default L4 GPU:

modal deploy comfyui.py

Or rebuild the image and deploy (ie. after removing models at persistent volume to redownload them):

MODAL_FORCE_BUILD=1 modal deploy comfyui.py

Or deploy with cleared shared_dict (ie. when the App forcefully stopped):

python comfyui.py

Or change the GPU with:

MODAL_GPU=RTX-PRO-6000 modal deploy comfyui.py

You can find the GPU types available on modal.com at https://modal.com/docs/guide/gpu

Other Environment Variables you can use are:

COMFY_VER="latest" # or "nightly --commit <hash>" or "nightly --pr kijai:minimax_fun"
COMFYGPU_ARGS="--use-flash-attention --preview-method auto --front-end-version Comfy-Org/ComfyUI_frontend@1.45.21"
COMFYMIX_ARGS="--preview-method auto --front-end-version Comfy-Org/ComfyUI_frontend@1.45.21"
JOBS_CUTOFFTIME=172800
MODAL_MAXTIME=3600
MODAL_IDLETIME=38
MODAL_WAITTIME=20
MODAL_MAXSTARTTIME=300

You can access ComfyUI from the provided persistent URL when successfully deployed.

Features

• Auto-scaling: Scales down to zero when not in use to save costs (modal's serverless can also auto-scales vertically, where CPU cores and RAM size can grow automatically as needed, so you don't need to overprovision them).
• GPU Snapshots: Fast startup times using Modal's GPU snapshots (cold-start can be under 3 seconds).
• Model Caching: Uses Modal Volumes to cache models across runs (modal's persistent volume is free for the first 1 TiB).
• Custom Node Management: Integrated with comfy-cli for easy plugin installation.
• Mixed CPU and GPU instance: Create/Edit your workflows using CPU-only instance for cheaper rates, but runs workflows on GPU instance seamlessly. Also have persistent completed jobs across sessions with their output assets accessible from Media Assets panel.
• Pre-installed Wheels:
  • PyTorch+CUDA 13.0
  • FlashAttention 2.8.3, 3, and 4
  • SageAttention 2.2 and 3
  • SpargeAttention 0.1
  • nunchaku 1.2.1
  • llama-cpp-python

Contributing

Please feel free to contribute to make this project better. Performance improvements/optimizations are very welcome.