In this guide you will learn:

• →Set up a clean local development environment using standard repository paths
• →Build and bundle complex workflows for deployment
• →Deploy ComfyUI to cloud providers like RunPod, Modal, and Replicate
• →Integrate CI/CD pipelines with GitHub Actions for automated updates
• →Secure and monitor production-level AI endpoints

ComfyUI Deployment Guide: From Local to Production

Deploying ComfyUI isn't just about running a script—it's about building a robust, repeatable infrastructure for AI generation. Whether you're building a private creative studio or a public-facing API, this guide covers the engineering required to move from experimental workflows to production power.

:::stats :::stat Configuration-dependent | Hardware planning :::stat 18 min | Reading Time :::stat Docker | Scale Method :::stat Production | Readiness :::

#1 — Overview

ComfyUI is the most efficient interface for professional AI generation due to its stateless execution model. Unlike heavy UIs, ComfyUI can be run as a headless API, allowing you to trigger complex workflows (Image, Video, Audio) via JSON requests.

Benefits of Deployment

• →Scalability: Run multiple nodes to handle high batch volumes.
• →Portability: Move workflows between local machines and cloud GPUs in seconds.
• →Independence: Decouple the creative process (workflow building) from the inference process (scaling).

#2 — Prerequisites

Before deploying, ensure your environment matches the production target:

• →Python 3.10.x: Essential for custom node compatibility.
• →NVIDIA GPU Drivers: Latest CUDA 12.1+ supported drivers.
• →Git: For version control and dependency management.
• →Repo Structure: We recommend the following standard layout (available in this repository):

#3 — Local Installation

Standardize your local setup to ensure "it just works" when moving to the cloud.

:::tip Use Symbolic Links For large model files, use mklink (Windows) or ln -s (Linux) to point models/ to a shared drive. This prevents duplicating multi-GB checkpoints across your repo. :::

#4 — Custom Node Installation

Most workflows require custom nodes. We recommend managing these specifically via the custom_nodes/ folder.

1. →Manual Install: Clone the node repo directly into custom_nodes/.
2. →Manager Install: Use the ComfyUI Manager (cloned into custom_nodes/) for one-click updates.

#5 — Packaging Workflows

To deploy a workflow, you must export it as an API-formatted JSON.

1. →In ComfyUI, enable "Dev mode" in settings.
2. →Click "Save (API Format)".
3. →Store this file in /workflows/ in your repository.

:::warning Regular Save vs API Format A standard "Save" includes UI metadata like node positions. The "API Format" removes this, making it significantly smaller and faster for headless execution. :::

#6 — Deployment Options

Local Server

Run with --listen to accept requests from your local network.

Cloud Deployment

• →Docker: Containerize the entire environment for Kubernetes or AWS.
• →RunPod / Lambda: Best for manual scaling and long-running GPU sessions.
• →Modal: Ideal for "serverless" deployments—pay only for the seconds you generate.

#7 — Running Production APIs

Once deployed, interact with ComfyUI via the /prompt endpoint. Your request body should contain the JSON from /workflows/.

:::note Securing Access Always run ComfyUI behind a reverse proxy (like Nginx) or a VPN. Do not expose port 8188 directly to the public internet. :::

#8 — CI/CD Integration

Use GitHub Actions to keep your production server synced with your repository.

Example Pipeline Logic:

1. →Push new model to models/ or workflow to workflows/.
2. →GitHub Action triggers a webhook.
3. →Server executes a git pull and restarts the ComfyUI service.

#9 — Monitoring & Logging

Production deployments require visibility into GPU usage and error rates.

• →NVIDIA SMI: Monitor VRAM usage (nvidia-smi -l 1).
• →Standard Logs: Capture stdout to a file in /scripts/logs/.
• →Latency Tracking: Measure the time between /prompt submission and file output.

#10 — Troubleshooting

:::warning Avoid "OOM" Errors If your API returns a 500 error during generation, check the console for "Out of Memory". Reduce your resolution or use a lower-bit quantized model (GGUF). :::

Common Node Errors:

• →ImportError: Ensure you've run pip install -r requirements.txt inside the specific custom node's folder.
• →Node not found: Verify the folder name in custom_nodes/ matches exactly.

#11 — Next Steps

• →Add Models: Populate /models/checkpoints/ with FLUX or SDXL models.
• →Share: Use /scripts/export_env.sh to share your exact environment with teammates.
• →Scale: Integrate with our Hardware Estimator to find the best cloud GPU for your budget.

#References

• →Official ComfyUI Docs
• →BlackMORE Ops Setup Guide
• →DiffHub Installation
• →ComfyUI-Deploy Extension
• →RunPod Blog