XPOSED

XPOSED

@XPOSED

Created:January 20, 2026 at 02:52 AM
Last edited:January 23, 2026 at 06:39 AM

How to Run ComfyUI on RunPod (Beginner Friendly)

This guide shows you how to run ComfyUI in the cloud using RunPod and generate high-quality images with the CyberRealistic Pony XL model and five LoRAs.

No local GPU required.

You’ll be able to follow this even if you’ve never used RunPod before.



1. Create a RunPod Account (Optional Bonus)

https://runpod.io

If you have a Google account, RunPod may show a modal with a “Claim Bonus” button.

Uploaded image

  • Click Claim Bonus
  • You’ll be redirected to the signup page
  • Sign up with Google to receive the bonus

    • Signing up without Google works too, but without the bonus

    

    Top Up Your Balance After signing in:

    Uploaded image
    
    1. Go to Billing
    2. Top up your balance
    • Minimum: $10
    • Payment methods: Card or crypto

    This balance is used to pay for GPU time.

    

    Create a New Pod (Choose the Right GPU)

    Navigate to Pods in the RunPod sidebar.

    You’ll see the pod configuration screen where you select your GPU.

    Uploaded image
    

    GPU Selection (Important)

    For this tutorial, you have two recommended options, depending on your experience level:

    

    🟢 RTX 2000 Ada (Recommended for First Experiments)

    • Lower cost
    • Fully capable of running:
    1. CyberRealistic Pony XL
    2. 5 LoRAs
    3. 832×1216 image generation
    • Slightly slower, but very stable

    👉 Choose this if you’re new to ComfyUI or just testing things out.

    

    🔵 RTX 4090 (For Fast & Confident Workflows)

    • Much faster generation
    • Better for:
    1. Rapid iteration
    2. Higher resolutions
    3. Batch generation
    4. More LoRAs and models used in workflow.

    👉 Choose this if you already know what you’re doing and want speed.

    

    💡 Note

    Both GPUs work perfectly for this guide.

    The only difference is generation speed and cost, not image quality.

    Once selected, continue with the rest of the setup exactly the same.

    

    Choose the ComfyUI Template Scroll down to Templates and click Change Template.

    Uploaded image
    

    In the modal:

    Uploaded image
    
    1. Search for ComfyUI
    2. Select: “ComfyUI”

    

    Edit Template Settings (Important)

    Uploaded image
    

    Click Edit Template, then adjust the following:

    Disk Configuration

    • Container Disk: 50 GB
    • Volume Disk: 70 GB

    This gives enough space for:

    • Checkpoints
    • LoRAs
    • WAN models
    • Outputs

    Environment Variables

    Find:

    CIVITAI_TOKEN

    Replace it with your own CivitAI API token.

    How to Get Your CivitAI Token

    1. https://civitai.com/user/account
    2. Scroll down until you find API keys.
    3. Create new key by clicking "Add API key" and copy it.

    Paste this token into the CIVITAI_TOKEN field in RunPod.

    This allows automated model downloads.

    Click on "Set Overrides"

    Uploaded image
    

    Deploy the Pod After setting everything:

    Uploaded image
    

    Click Deploy On-Demand Your pod will begin provisioning. You can open Logs to watch:

    • CUDA setup
    • ComfyUI installation

    Wait until the pod status shows Running.

    

    9. Enable web terminal and Download Models

    Uploaded image
    

    Enable web terminal and open it.

    

    Download Models Using a Script

    Create and run a download script to automatically fetch your models:

    Step 1: Create the script file

    nano download_models.sh

    Step 2: Paste this script:

    #!/bin/bash

# ComfyUI Model Download Script
# This script downloads AI image generation models for ComfyUI

# Set base directory for ComfyUI models
BASE_DIR="/workspace/runpod-slim/ComfyUI/models"

# Create directory structure
mkdir -p "$BASE_DIR/checkpoints" "$BASE_DIR/loras"

echo "Starting model downloads..."

# ============================================
# CHECKPOINT MODELS
# ============================================
# Checkpoints are the main AI models that generate images.
# They contain the core neural network weights trained on millions of images.
# These are typically 2-7GB files that define the base style and capabilities.

echo "Downloading checkpoint model..."
wget -O "$BASE_DIR/checkpoints/cyberrealisticPony_v150.safetensors" \
  "https://civitai.com/api/download/models/2469412?token=${CIVITAI_TOKEN}"

# ============================================
# LORA MODELS (Low-Rank Adaptation)
# ============================================
# LoRAs are small supplementary models (50-200MB) that modify the checkpoint.
# They add specific styles, concepts, or characteristics without retraining
# the entire base model. You can use multiple LoRAs together.

echo "Downloading LoRA models..."

# Realistic skin texture enhancement
wget -O "$BASE_DIR/loras/RealSkin_xxXL_v1.safetensors" \
  "https://civitai.com/api/download/models/1681921?token=${CIVITAI_TOKEN}"

# Raw camera/photography style
wget -O "$BASE_DIR/loras/RawCam_250_v1.safetensors" \
  "https://civitai.com/api/download/models/1926656?token=${CIVITAI_TOKEN}"

# Dramatic lighting effects control
wget -O "$BASE_DIR/loras/Dramatic Lighting Slider.safetensors" \
  "https://civitai.com/api/download/models/1242203?token=${CIVITAI_TOKEN}"

# Leaked nudes aesthetic style
wget -O "$BASE_DIR/loras/leaked_nudes_style_v1_fixed.safetensors" \
  "https://civitai.com/api/download/models/1627770?token=${CIVITAI_TOKEN}"

# Amateur photography style
wget -O "$BASE_DIR/loras/amateur_style_v1_pony.safetensors" \
  "https://civitai.com/api/download/models/1594293?token=${CIVITAI_TOKEN}"

echo "Download complete!"
echo "Checkpoint saved to: $BASE_DIR/checkpoints/"
echo "LoRAs saved to: $BASE_DIR/loras/"#!/bin/bash

# ComfyUI Model Download Script
# This script downloads AI image generation models for ComfyUI

# Set base directory for ComfyUI models
BASE_DIR="/workspace/runpod-slim/ComfyUI/models"

# Create directory structure
mkdir -p "$BASE_DIR/checkpoints" "$BASE_DIR/loras"

echo "Starting model downloads..."

# ============================================
# CHECKPOINT MODELS
# ============================================
# Checkpoints are the main AI models that generate images.
# They contain the core neural network weights trained on millions of images.
# These are typically 2-7GB files that define the base style and capabilities.

echo "Downloading checkpoint model..."
wget -O "$BASE_DIR/checkpoints/cyberrealisticPony_v150.safetensors" \
  "https://civitai.com/api/download/models/2469412?token=${CIVITAI_TOKEN}"

# ============================================
# LORA MODELS (Low-Rank Adaptation)
# ============================================
# LoRAs are small supplementary models (50-200MB) that modify the checkpoint.
# They add specific styles, concepts, or characteristics without retraining
# the entire base model. You can use multiple LoRAs together.

echo "Downloading LoRA models..."

# Realistic skin texture enhancement
wget -O "$BASE_DIR/loras/RealSkin_xxXL_v1.safetensors" \
  "https://civitai.com/api/download/models/1681921?token=${CIVITAI_TOKEN}"

# Raw camera/photography style
wget -O "$BASE_DIR/loras/RawCam_250_v1.safetensors" \
  "https://civitai.com/api/download/models/1926656?token=${CIVITAI_TOKEN}"

# Dramatic lighting effects control
wget -O "$BASE_DIR/loras/Dramatic Lighting Slider.safetensors" \
  "https://civitai.com/api/download/models/1242203?token=${CIVITAI_TOKEN}"

# Leaked nudes aesthetic style
wget -O "$BASE_DIR/loras/leaked_nudes_style_v1_fixed.safetensors" \
  "https://civitai.com/api/download/models/1627770?token=${CIVITAI_TOKEN}"

# Amateur photography style
wget -O "$BASE_DIR/loras/amateur_style_v1_pony.safetensors" \
  "https://civitai.com/api/download/models/1594293?token=${CIVITAI_TOKEN}"

echo "Download complete!"
echo "Checkpoint saved to: $BASE_DIR/checkpoints/"
echo "LoRAs saved to: $BASE_DIR/loras/"
    python

    Step 3: Save and exit

    1. Press Ctrl + O to save
    2. Press Enter to confirm
    3. Press Ctrl + X to exit

    Step 4: Make the script executable

    chmod +x download_models.sh

    Step 5: Run the script

    ./download_models.sh

    The script will download all models to the correct directories. You can monitor the progress in the terminal.

    

    10. Open ComfyUI

    Go to the Connect tab on RunPod.

    Uploaded image
    

    Click the address that contains:

    :8188

    This opens ComfyUI in your browser.

    

    11. Load the Image Generation Workflow

    Uploaded image
    

    This image is a ComfyUI workflow.

    Steps:

    1. Download it.
    2. Drag & drop the image into ComfyUI
    3. Open Node Manager
    4. Install any missing nodes
      5. Uploaded image
      
    5. Restart ComfyUI
      6. Uploaded image
      

      

    12. Understanding the Workflow (Before You Generate)

    The workflow attached to this article already contains:

    • A fixed seed
    • A pre-filled prompt
    • Final sampler settings

    This means that if you run it as-is, ComfyUI will generate the exact same image every time.

    This is intentional, so you can verify that everything is set up correctly.

    Once you confirm it works, you’re expected to experiment with prompt.

    

    13. Model & Sampler Settings Explained

    This workflow uses CyberRealistic Pony XL, which is based on PonyXL.

    Because of that, it works best with Danbooru-style tags instead of natural language.

    You can reference Danbooru tags here:

    

    Recommended Sampler Settings for This Model

    CyberRealistic Pony XL performs best with the following:

    Sampling method:

    • DPM++ SDE Karras
    • DPM++ 2M Karras
    • Euler a

    Sampling steps:

    • 30+ steps recommended

    Resolution (portrait works best):

    • 896×1152
    • 832×1216

    

    Settings Used in This Workflow

    The attached workflow is configured as follows:

    • Resolution: 832×1216
    • Sampler: DPM++ SDE Karras
    • Steps: 30
    • CFG: 5

    These values are a solid baseline for realistic Pony-based generations.

    

    14. Prompt Structure (Very Important)

    Pony-based models support score tokens, which strongly improve output quality and consistency.

    The workflow uses this structure:

    score_9, score_8_up, score_7_up, (SUBJECT),score_9, score_8_up, score_7_up, (SUBJECT),
    plaintext

    Why scores matter

    • They bias the model toward higher-quality training samples
    • Improve anatomy, lighting, and composition
    • Reduce randomness without increasing CFG

    This is one of the main reasons Pony-based models feel more stable at lower CFG values.

    

    Writing Your Own Prompt

    Replace (SUBJECT) with Danbooru-style tags, for example:

    score_9, score_8_up, score_7_up, 1girl, solo, long hair, realistic lighting, detailed face, doggy style, closeupscore_9, score_8_up, score_7_up, 1girl, solo, long hair, realistic lighting, detailed face, doggy style, closeup
    plaintext
    Uploaded image
    

    You can:

    • Add environment tags
    • Add clothing tags
    • Add camera or lighting tags

    Avoid full sentences.

    Tags > prose for this model.

    

    15. Generating Your Own Image

    After editing the prompt:

    1. Keep the sampler and resolution as-is
    2. Click Queue Prompt
    3. Let it generate

    If you want variation and experiment you may:

    • Change seed value (you can also set it to randomize and it will be different on every run)
    • Slightly adjust CFG (4.5–6)
    • Adjust LoRA weights

    

    Final Tip

    This workflow is meant to be a clean, beginner-friendly, reproducible baseline.

    Once you understand how it behaves, you can:

    • Swap LoRAs
    • Change resolutions
    • Try different samplers
    • Build your own workflows

    Comments

    No comments yet. Be the first to comment!