ComfyUI on a VPS: Self-Hosted Stable Diffusion Image Generation with Node-Based Workflow

By VPS.DO
September 15, 2026

ComfyUI is the most powerful and flexible open-source Stable Diffusion interface — it uses a node-based visual workflow editor where each step of the image generation pipeline (model loading, conditioning, sampling, upscaling, face restoration) is a separate node that you connect visually. Unlike Automatic1111 (linear UI), ComfyUI’s workflow approach enables complex pipelines, reproducible generation workflows, and a JSON-based workflow format that works as an API. Self-hosting eliminates Midjourney ($10–$30/month) and Stable Diffusion API fees ($0.003–$0.02/image).

ComfyUI vs Automatic1111

ComfyUI: Node-based workflow editor, more flexible for complex pipelines, better performance, workflow JSON exports, built-in API. Learning curve is steeper.
Automatic1111: Traditional web UI, more user-friendly for beginners, huge extension ecosystem. Not as performant.
Choose ComfyUI: Reproducible workflows, API integration, complex multi-step pipelines (img2img + upscale + face restore in one workflow).

Server Requirements

GPU (recommended): NVIDIA GPU with 6+ GB VRAM for SDXL, 4 GB for SD1.5. CUDA 12.1+
CPU (functional, slow): 16+ GB RAM, expect 1–5 minutes per image vs 3–10 seconds on GPU
50+ GB disk for models (SDXL ~6 GB, Flux ~24 GB)
Python 3.10+

Step 1: Install ComfyUI

<code"># Create environment
sudo apt install -y python3.11 python3.11-venv git
python3.11 -m venv /opt/comfyui-env
source /opt/comfyui-env/bin/activate

# Clone ComfyUI
git clone https://github.com/comfyanonymous/ComfyUI.git /opt/comfyui
cd /opt/comfyui

# Install dependencies
# For CPU only:
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
pip install -r requirements.txt

# For NVIDIA GPU:
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

Step 2: Download Models

<code"># Model storage locations:
# /opt/comfyui/models/checkpoints/  — main SD models
# /opt/comfyui/models/loras/        — LoRA fine-tunes
# /opt/comfyui/models/vae/          — VAE models
# /opt/comfyui/models/upscale_models/ — ESRGAN upscalers

# Download SDXL base model (6.5 GB)
pip install huggingface_hub
huggingface-cli download \
    stabilityai/stable-diffusion-xl-base-1.0 \
    sd_xl_base_1.0.safetensors \
    --local-dir /opt/comfyui/models/checkpoints/

# Download Flux model (more powerful, 23 GB)
# huggingface-cli download black-forest-labs/FLUX.1-schnell \
#     --local-dir /opt/comfyui/models/checkpoints/

# Download VAE (improves color accuracy)
wget -O /opt/comfyui/models/vae/sdxl_vae.safetensors \
    https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/resolve/main/sdxl_vae.safetensors

Step 3: Start ComfyUI

<code"># CPU mode
cd /opt/comfyui
source /opt/comfyui-env/bin/activate
python main.py \
    --listen 127.0.0.1 \
    --port 8188 \
    --cpu                   # CPU inference mode
    # --gpu-only            # GPU mode (remove --cpu)
    # --lowvram             # For GPU with limited VRAM

Step 4: Systemd Service

<code">sudo nano /etc/systemd/system/comfyui.service

<code">[Unit]
Description=ComfyUI Stable Diffusion
After=network.target

[Service]
Type=simple
User=ubuntu
WorkingDirectory=/opt/comfyui
ExecStart=/opt/comfyui-env/bin/python main.py --listen 127.0.0.1 --port 8188 --cpu
Restart=on-failure
RestartSec=10

[Install]
WantedBy=multi-user.target

<code">sudo systemctl daemon-reload
sudo systemctl enable comfyui
sudo systemctl start comfyui

Step 5: Nginx with Authentication

<code">sudo nano /etc/nginx/sites-available/comfyui

<code">server {
    listen 443 ssl http2;
    server_name diffusion.yourdomain.com;

    ssl_certificate /etc/letsencrypt/live/diffusion.yourdomain.com/fullchain.pem;
    ssl_certificate_key /etc/letsencrypt/live/diffusion.yourdomain.com/privkey.pem;

    # Basic auth — protect your GPU resource
    auth_basic "ComfyUI";
    auth_basic_user_file /etc/nginx/.comfyui_htpasswd;

    client_max_body_size 100M;

    location / {
        proxy_pass http://127.0.0.1:8188;
        proxy_http_version 1.1;
        proxy_set_header Upgrade $http_upgrade;
        proxy_set_header Connection 'upgrade';
        proxy_set_header Host $host;
        proxy_read_timeout 600s;  # Long timeout for image generation
        proxy_buffering off;
    }
}

<code">sudo htpasswd -c /etc/nginx/.comfyui_htpasswd admin
sudo certbot --nginx -d diffusion.yourdomain.com
sudo systemctl reload nginx

Step 6: API Usage for Application Integration

<code">import json
import httpx
import uuid
import time

# Export a workflow from ComfyUI UI → Save → download JSON
# Then use it in API calls:

with open('my_workflow.json') as f:
    workflow = json.load(f)

# Modify workflow parameters programmatically
workflow['6']['inputs']['text'] = "a majestic mountain at sunset, photorealistic"
workflow['7']['inputs']['text'] = "blurry, low quality, watermark"
workflow['3']['inputs']['seed'] = random.randint(0, 2**32)

# Queue the generation
client_id = str(uuid.uuid4())
response = httpx.post('http://localhost:8188/prompt', json={
    'prompt': workflow,
    'client_id': client_id,
})
prompt_id = response.json()['prompt_id']

# Poll for completion
while True:
    history = httpx.get(f'http://localhost:8188/history/{prompt_id}').json()
    if prompt_id in history:
        outputs = history[prompt_id]['outputs']
        # Get the image filename
        for node_id, output in outputs.items():
            if 'images' in output:
                for img in output['images']:
                    filename = img['filename']
                    # Download the generated image
                    image_data = httpx.get(
                        f'http://localhost:8188/view?filename={filename}&type=output'
                    ).content
                    with open(f'output_{filename}', 'wb') as f:
                        f.write(image_data)
                    print(f"Generated: {filename}")
        break
    time.sleep(2)

Step 7: Custom Nodes (Extend ComfyUI)

<code"># Install ComfyUI Manager for easy custom node installation
cd /opt/comfyui/custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Manager.git

# Popular custom nodes:
# ComfyUI-Impact-Pack: face detection, seam fixing, detailers
# ComfyUI_IPAdapter_plus: style transfer from reference images
# ComfyUI-VideoHelperSuite: video generation workflows
# ComfyUI-AnimateDiff: animated image generation

pip install -r ComfyUI-Manager/requirements.txt
sudo systemctl restart comfyui

Getting Started

For GPU inference, a VPS with NVIDIA A4000 (16 GB VRAM) generates SDXL images in 3–10 seconds. For CPU-only inference on a high-RAM Ubuntu VPS at VPS.DO (16+ GB RAM), expect 1–5 minutes per SDXL image — adequate for batch generation workflows where speed isn’t critical. NVMe storage speeds up model loading significantly — SDXL takes 10 seconds to load from NVMe versus 60+ seconds from HDD.

Conclusion

ComfyUI on a self-hosted VPS provides unlimited AI image generation with reproducible node-based workflows, a JSON API for application integration, and no per-image fees. For developers building image generation into applications (product mockups, avatar generation, marketing assets), ComfyUI’s JSON workflow API is the cleanest interface for programmatic generation. GPU-equipped VPS instances make generation speeds comparable to cloud APIs at a fraction of the per-image cost at scale.

ComfyUI on a VPS: Self-Hosted Stable Diffusion Image Generation with Node-Based Workflow