ComfyUI on a Cloud GPU: Run Stable Diffusion and FLUX Remotely (2026)

Contains affiliate links. Disclosure.

ComfyUI is a node-based workflow editor for Stable Diffusion and FLUX. It is more powerful and flexible than AUTOMATIC1111, with a steeper learning curve. The catch: SDXL wants 12GB+ VRAM, and FLUX wants 24GB+. Most consumer GPUs fall short. The solution: run ComfyUI on a cloud GPU server and access it via SSH tunnel or reverse proxy. This guide covers setup on a fresh GPU server, model management, essential workflows, and remote access. For hosting, Liquid Web offers GPU VPS with RTX 4090, A100, and H100 options.

What is ComfyUI? A node-based Stable Diffusion interface where you wire loaders, samplers, and output nodes into custom pipelines. You can build multi-step workflows (inpaint → upscale → refine) that AUTOMATIC1111 cannot express. Perfect for production image pipelines and automation. See ComfyUI + n8n to trigger generation from workflows.

Get a Liquid Web GPU server → | Apify for image data pipelines →

What Is ComfyUI and Why Cloud GPU?

ComfyUI uses a node-based graph instead of a single form. You connect nodes for loading models, setting samplers, applying LoRAs, and saving outputs. This enables complex workflows that would require scripting in AUTOMATIC1111.

Why cloud GPU?

• SDXL needs ~12 GB VRAM. FLUX.1 needs ~24 GB for comfortable generation.
• Consumer GPUs (RTX 3060, 3070) often lack enough VRAM or slow down with large batch sizes.
• A rented GPU server gives you RTX 4090, A100, or A10G on demand — no upfront hardware cost.

Liquid Web GPU VPS provides these tiers with hourly or monthly billing.

---

Setup on a Fresh GPU Server

Prerequisites: Ubuntu 22.04, NVIDIA driver installed, nvidia-smi working.

# Verify GPU
nvidia-smi

1. Clone ComfyUI

git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI

2. Create virtual environment

python3 -m venv venv
source venv/bin/activate

3. Install dependencies

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

For FLUX and SDXL, the default requirements are usually sufficient. ComfyUI Manager adds extra nodes — we'll cover that next.

4. Run ComfyUI with remote access

python main.py --listen

--listen binds to 0.0.0.0, so ComfyUI is reachable from outside localhost. Default port is 8188.

---

SSH Tunnel for Secure Access

Exposing ComfyUI directly to the internet can be risky. Use an SSH tunnel:

ssh -L 8188:localhost:8188 user@YOUR_SERVER_IP

Keep this SSH session open. On your local machine, open http://localhost:8188. All traffic goes through the encrypted SSH tunnel.

---

Essential Nodes and ComfyUI Manager

ComfyUI Manager lets you install custom nodes from a catalog. It's the first add-on to install:

1. Open ComfyUI in the browser.
2. Click Manager (or install via git: cd custom_nodes && git clone https://github.com/ltdrdata/ComfyUI-Manager.git).
3. Restart ComfyUI.
4. Use Manager to search and install nodes (e.g. ControlNet, upscalers).

Common nodes: Load Checkpoint, KSampler, VAE Decode, Save Image, Load LoRA.

---

FLUX.1 Workflow Example

A minimal FLUX workflow:

1. Load Checkpoint — Point to your FLUX checkpoint (flux1-dev.safetensors or similar) in models/checkpoints.
2. CLIP Text Encode (Prompt) — Your positive prompt.
3. CLIP Text Encode (Negative) — Negative prompt (optional).
4. Empty Latent Image — Width 1024, height 1024 (or 768 for SD 1.5).
5. KSampler — Steps 28–35, cfg 3.5–5, sampler euler or dpmpp_2m.
6. VAE Decode — Connect KSampler output.
7. Save Image — Output directory and filename.

Wire: Checkpoint → KSampler; Prompt → KSampler; Empty Latent → KSampler → VAE Decode → Save Image.

---

SDXL Workflow with LoRA

1. Load Checkpoint — SDXL base (e.g. sd_xl_base_1.0.safetensors).
2. Load LoRA — Select LoRA file from models/loras. Set strength (0.7–0.9 typical).
3. CLIP Text Encode — Prompt. SDXL supports longer prompts.
4. Empty Latent Image — 1024×1024 for SDXL.
5. KSampler — Steps 25–30, cfg 7–8.
6. VAE Decode → Save Image.

Place LoRA files (.safetensors) in ComfyUI/models/loras/. For API-driven automation, ComfyUI + n8n lets you trigger workflows from webhooks without touching the UI.

---

Port Forwarding and Reverse Proxy

For team access without SSH tunnels, expose ComfyUI via a reverse proxy:

1. Nginx — Proxy https://your-domain.com to http://localhost:8188. Add Basic Auth or OAuth if needed.
2. Traefik — Use Traefik v3 for automatic HTTPS and routing.
3. Cloudflare Tunnel — Free HTTPS tunnel without opening ports; useful for quick sharing.

Never expose ComfyUI directly to the internet without authentication — generation can consume significant GPU time and attract abuse. The same Liquid Web server that runs ComfyUI can also host vLLM for text generation if you need both modalities.

---

Model Management

Type	Directory	Source
Checkpoints	models/checkpoints/	HuggingFace, CivitAI
LoRAs	models/loras/	CivitAI
VAE	models/vae/	Often bundled with checkpoint
ControlNet	models/controlnet/	HuggingFace

Download from HuggingFace:

cd ComfyUI/models/checkpoints
wget https://huggingface.co/.../model.safetensors

ComfyUI Manager can download from CivitAI and HuggingFace via the UI. Keep checkpoints organized by use case: one folder for SDXL base models, another for FLUX, and separate LoRAs for styles (portrait, landscape, anime). This speeds up workflow switching when you automate with n8n.

---

ComfyUI vs AUTOMATIC1111 vs InvokeAI

Feature	ComfyUI	AUTOMATIC1111	InvokeAI
Interface	Node-based graph	Form-based	Form + canvas
Workflow flexibility	High (custom pipelines)	Low	Medium
Learning curve	Steep	Gentle	Medium
API mode	Yes (/prompt endpoint)	Limited	Yes
FLUX support	Yes	Partial	Yes
Production automation	Best	Limited	Good

Choose ComfyUI when you need reproducible workflows, API-driven generation, or complex multi-step pipelines. The same Liquid Web GPU server can run vLLM for LLMs (see Serve LLaMA 3 with vLLM) and ComfyUI for images. To automate generation, connect ComfyUI with n8n; for custom models, fine-tune LLaMA 3 with QLoRA before serving.

Start with SDXL

FLUX produces higher quality but needs 24GB VRAM. If you have a 12GB GPU, start with SDXL. Once you're comfortable with nodes, add FLUX when you upgrade. Liquid Web offers both 12GB and 24GB+ options.

Frequently Asked Questions

SDXL: 12GB VRAM minimum (RTX 3060 12GB, 4060 Ti 16GB). FLUX: 24GB recommended (RTX 4090, A100 40GB). For batch generation, more VRAM helps.

SSH tunnel: ssh -L 8188:localhost:8188 user@server. Then open http://localhost:8188. For production, use a reverse proxy (Nginx/Caddy) with HTTPS and optionally auth.

Yes. Run headless on a server. Use --listen for network access. No X11 or display needed.

In ComfyUI/models/checkpoints/. Download flux1-dev.safetensors or similar from HuggingFace. ComfyUI Manager can also fetch models.

ComfyUI: node-based, more powerful, API-friendly, steeper curve. AUTOMATIC1111: simpler, form-based, easier for beginners. ComfyUI is better for automation and production pipelines.