Run OpenWebUI + ComfyUI

Combine Open WebUI + Ollama + ComfyUI in a single GPU workload to use Korean-language chat and image generation together.

Before You Begin

Item	Description
gcube account	Sign up at gcube.ai
Point balance	Billed hourly based on GPU usage time

Step 1 — Register Workload

1-1. Register New Workload

Go to the workload page and register a new workload or modify an existing one.

New workload registration screen

1-2. Container Settings

Enter the following information.

Item	Value
Registry Type	GitHub
Container Image	`open-webui/open-webui:ollama`
Container Port	Auto-filled after image validation

Tip

After entering the registry type and container image, click the Validate button. If the image is valid, the container port will be filled in automatically.

Container settings screen

1-3. Select GPU and Register

Select a GPU and complete registration with Manual Deployment.

GPU selection screen

1-4. Deploy Workload

Click the Deploy button on the workload management screen.

Workload deploy button

Once deployment starts, you can check the service URL activation and deployment progress.

Deployment progress screen

Warning

Deployment may take 5–10 minutes depending on the container image size and network environment. For Tier 1 cloud GPUs, it may take up to 30 minutes.

Step 2 — Access Open WebUI and Create Admin Account

2-1. Access Open WebUI

Click the service URL to access the Open WebUI web page.

Service URL click screen

2-2. Create Admin Account

Create an admin account on first access.

Admin account creation screen

2-3. Check LLM Model

After accessing Open WebUI, you can select an LLM model to use conversational AI. However, since no LLM has been downloaded yet, it cannot be used at this stage.

In the next step, you will install the Bllossom model based on Llama-3, which has strong Korean language capabilities.

LLM model selection screen

Step 3 — Install Korean LLM (Bllossom)

3-1. Access Workload Terminal

Click the running workload title to go to the workload info screen.

Workload info screen

Click the Container Terminal button to open the terminal.

Container terminal access screen

3-2. Download Bllossom GGUF Model

Enter the following commands in the terminal in order.

# Create Bllossom GGUF download folder
mkdir -p /root/models/bllossom
cd /root/models/bllossom

# Install Hugging Face download tool
python3 -m pip install -U huggingface_hub

# Download GGUF model
python3 -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='MLP-KTLim/llama-3-Korean-Bllossom-8B-gguf-Q4_K_M', local_dir='.', local_dir_use_symlinks=False)"

Bllossom download screen

3-3. Create Modelfile

cat > /root/models/bllossom/Modelfile <<'EOF'
FROM ./llama-3-Korean-Bllossom-8B-Q4_K_M.gguf

SYSTEM """
You are a helpful AI assistant.
Please answer the user's questions kindly and accurately.
The default response language is Korean.
"""
EOF

3-4. Register and Run Model in Ollama

# Register model in Ollama
cd /root/models/bllossom
ollama create korean-bllossom -f Modelfile

# Run
ollama run korean-bllossom

Once registration is complete, you can use the Bllossom model for Korean conversational AI in Open WebUI.

Bllossom Korean chat screen

Step 4 — Install ComfyUI and Download Image Generation Model

4-1. Install ComfyUI

Enter the following commands in the gcube workload terminal in order.

# Install system packages
apt update && apt install -y git python3-venv python3-pip

# Create ComfyUI folder
mkdir -p /comfyui
cd /comfyui

# Download source
git clone https://github.com/comfy-org/ComfyUI.git
cd /comfyui/ComfyUI

# Create and activate virtual environment
python3 -m venv .venv
. .venv/bin/activate

# Upgrade pip
pip install -U pip

# Install PyTorch for NVIDIA (cu118)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

# Install ComfyUI dependencies
pip install -r requirements.txt

# Run in background
nohup python3 main.py --listen 0.0.0.0 > /comfyui/comfyui.log 2>&1 &

# Check logs
tail -f /comfyui/comfyui.log

4-2. Download ByteDance/SDXL-Lightning Model

# Navigate to ComfyUI folder
cd /comfyui/ComfyUI

# Activate virtual environment
[ -f .venv/bin/activate ] && . .venv/bin/activate

# Create checkpoints folder
mkdir -p models/checkpoints

# Install download tools
python3 -m pip install -U pip huggingface_hub hf_xet

# Download SDXL-Lightning 4step model
HF_XET_HIGH_PERFORMANCE=1 hf download ByteDance/SDXL-Lightning sdxl_lightning_4step.safetensors --local-dir ./models/checkpoints

Step 5 — Configure Open WebUI ↔ ComfyUI Integration

5-1. Access Admin Panel

Access the Open WebUI admin panel.

Admin panel access screen

5-2. Enter Image Generation Settings

Go to Admin Panel → Settings → Images and enter the following.

Item	Value
Image Generation	Enable (check)
Image Generation Engine	ComfyUI
ComfyUI Base URL	`http://127.0.0.1:8188`
Model	`sdxl_lightning_4step.safetensors`
Image Size	1024 × 1024
Steps	4

ComfyUI Workflow Settings:

Item	Value
Steps	4
CFG	1
Sampler	euler
Scheduler	sgm_uniform
ckpt_name	sdxl_lightning_4step.safetensors

ComfyUI Workflow Nodes:

Node	Value
text	6
ckpt_name	4
width	5
height	5
steps	3
seed	3

Image generation settings screen

ComfyUI workflow settings screen

5-3. Text Chat and Image Generation

After configuration, you can use text chat and image generation together in the Open WebUI chat window.

Open WebUI text and image generation screen