ComfyUI Local Docker Setup: Private Generative AI Without Cloud Telemetry (2026)

Executive Summary & Architecture Overview

Generative image AI platforms like Midjourney, DALL-E 3, and proprietary cloud services log every prompt, retain commercial usage rights, and censor technical outputs. For artists, developers, and homelabbers seeking zero telemetry, full privacy, and unlimited generation capabilities, ComfyUI has become the industry standard node-based execution engine for Stable Diffusion (SD 1.5, SDXL, Flux.1, and SD 3.5).

Unlike monolithic WebUIs, ComfyUI executes modular computation graphs with optimized VRAM paging, reducing memory overhead by up to 40% and supporting high-resolution image rendering on consumer GPUs.

+--------------------------------------------------------------------------+
|                     User Web Browser (Graph Editor UI)                   |
+------------------------------------^-------------------------------------+
                                     | HTTP / WebSocket (Port 8188)
+------------------------------------v-------------------------------------+
|                      ComfyUI Runtime Container                           |
|                      - PyTorch 2.4+ / CUDA 12.4 Runtime                  |
|                      - Dynamic VRAM Memory Paging                        |
|                      - Custom Nodes Manager (ComfyUI-Manager)            |
+------------------------------------^-------------------------------------+
                                     | Persistent Host Mounts
+------------------------------------v-------------------------------------+
|                     Host NVMe Storage Subsystem                          |
|                     - /models/checkpoints (Base Models: SDXL, Flux)      |
|                     - /models/loras (Style Adaptations)                  |
|                     - /models/vae & /models/clip                         |
|                     - /output (Rendered PNGs with embedded workflows)    |
+--------------------------------------------------------------------------+

Hardware & System Requirements

Step 1: Directory Setup & Host Model Paths

Organize model weights outside the container so you never lose downloaded checkpoints when updating images:

sudo mkdir -p /opt/comfyui/{models/checkpoints,models/loras,models/vae,models/clip,custom_nodes,output,input} && \
cd /opt/comfyui && \
sudo chown -R $USER:$USER /opt/comfyui

Step 2: Production-Grade docker compose.yaml

Save the following file as /opt/comfyui/compose.yaml (following modern Compose Spec V2 without version:):

services:
  comfyui:
    image: yanwk/comfyui-boot:cu124
    container_name: comfyui
    restart: unless-stopped
    ports:
      - "127.0.0.1:8188:8188"
    environment:
      - CLI_ARGS=--listen 0.0.0.0 --port 8188 --preview-method auto
    volumes:
      - /opt/comfyui/models:/root/ComfyUI/models
      - /opt/comfyui/custom_nodes:/root/ComfyUI/custom_nodes
      - /opt/comfyui/output:/root/ComfyUI/output
      - /opt/comfyui/input:/root/ComfyUI/input
    networks:
      - ai-network
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

networks:
  ai-network:
    name: ai-network
    driver: bridge

Start the container:

cd /opt/comfyui && docker compose up -d

Verify GPU visibility inside the runtime:

docker exec -it comfyui nvidia-smi

Step 3: Downloading Base Models & Workflows

Download a base SDXL model directly into your checkpoints folder:

cd /opt/comfyui/models/checkpoints && \
curl -L -O https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_base_1.0.safetensors

Install the essential ComfyUI-Manager to enable one-click custom node updates:

cd /opt/comfyui/custom_nodes && \
git clone https://github.com/ltdrdata/ComfyUI-Manager.git && \
docker compose restart comfyui

Access the interface at http://YOUR_SERVER_IP:8188 (or via reverse proxy).

Step 4: Reverse Proxy & Security

Never expose ComfyUI directly to the internet without authentication, as arbitrary custom Python nodes can execute arbitrary system code.

In Nginx Proxy Manager:

• Forward comfy.yourhomelab.net to http://127.0.0.1:8188.
• Enable Websockets Support.
• Under Access List, configure HTTP Basic Authentication or restrict access to your home subnet / Tailscale IP range.

Step 5: Remote Generation with Tailscale

Connect your mobile device or remote laptop to your Tailscale network and point your browser to http://100.x.y.z:8188. You can initiate complex image generation workflows from your phone while your homelab GPU handles the heavy tensor computations at home.

Step 6: Automated Workflow & Node Backup

Create /opt/comfyui/backup.sh:

#!/usr/bin/env bash
set -euo pipefail

BACKUP_DIR="/mnt/backups/comfyui"
DATE=$(date +"%Y%m%d_%H%M%S")

mkdir -p "\${BACKUP_DIR}"

# Backup custom node configurations and generated images metadata
tar -czf "\${BACKUP_DIR}/comfyui_config_\${DATE}.tar.gz" -C /opt/comfyui custom_nodes compose.yaml

find "\${BACKUP_DIR}" -type f -name "*.tar.gz" -mtime +14 -delete
echo "[✓] ComfyUI configuration backup completed."

Schedule weekly backups:

chmod +x /opt/comfyui/backup.sh && \
(crontab -l 2>/dev/null; echo "0 5 * * 0 /opt/comfyui/backup.sh >> /var/log/comfy-backup.log 2>&1") | crontab -

Step 7: Deep Troubleshooting Matrix

Step 8: Frequently Asked Questions (FAQ)

1. How does ComfyUI save workflows?

ComfyUI automatically embeds the entire visual node graph metadata into the metadata chunk of every generated PNG file. Dragging any generated image back into the web browser instantly restores the exact workflow, prompt, and seed.

2. Can I use AMD Radeon GPUs?

Yes. Substitute the NVIDIA image with the ROCm image (yanwk/comfyui-boot:rocm6.1) and pass /dev/kfd and /dev/dri devices in your compose file.

3. Where should I store LoRA weights?

Place all .safetensors LoRA files inside /opt/comfyui/models/loras/. They will immediately appear in the Load LoRA node dropdown without restarting.

4. What is the difference between SDXL and Flux.1?

SDXL is optimized for fast, photorealistic 1024x1024 generation requiring 8-12 GB VRAM. Flux.1 is a 12B parameter flow-matching model with state-of-the-art prompt fidelity and text rendering, requiring 16-24 GB VRAM.

5. How can I queue multiple renders remotely?

Use ComfyUI's queue feature or send API requests directly to its REST endpoint (/prompt). Workflows can be automated via Python scripts or Home Assistant webhooks.