Deploy ComfyUI on RunPod's cloud infrastructure for powerful AI image generation without local hardware requirements.
Deploy ComfyUI on RunPod's cloud infrastructure for powerful AI image generation without local hardware requirements.
By the end of this guide, you'll have:
If you're completely new to RunPod, we recommend starting with our Deploy your first Pod guide to learn the basics of creating an account and deploying your first GPU Pod. This guide will then show you how to specifically set up ComfyUI on your RunPod.
RunPod is a cloud GPU platform that lets you run ComfyUI on powerful hardware without investing in expensive local equipment. With on-demand access to high-end GPUs like RTX 4090s and L40s, you can handle complex workflows and large models that would be impossible on consumer hardware.
RunPod operates on a pay-per-use model:
Before starting, ensure you have:
This guide provides a comprehensive walkthrough of setting up and using RunPod for AI tasks, particularly with ComfyUI. It covers pod creation, template selection, GPU configuration, storage management, and essential tips for efficient workflow. This guide is based on the RunPod Tutorial 2025 video.
RunPod allows you to run AI models without needing expensive GPUs by renting computing power. The core component is the pod, which is an instance of a computer with memory and one or more GPUs.
Pods are essentially virtual machines with dedicated GPU resources.
Sign up on RunPod using the referral link provided by the video creator to support the channel.
Go to the "Pods" section in the RunPod interface.
Select a pod template from the "Hub" section under "Pod Templates". These templates pre-install operating systems and software like ComfyUI. Search for "Comfy UI basic endangered AI" to find templates created by the video creator.
Click "Deploy Pod" on the template page. Alternatively, use template links provided in the video descriptions or Discord channel.
Choose a GPU based on your VRAM requirements. Use the filter to find GPUs with the necessary VRAM. Be cautious when using multi-GPU setups unless your template is configured for it.
Newer GPUs (B200, H200, RTX 5090, RTX Pro 6000) may require updated CUDA versions, which older templates might not support.
Give your pod a name and verify the selected template. You can change the template if needed.
Choose between "On Demand" (pay-as-you-go) or reserved options (3, 6, or 12-month rentals for reduced hourly rates). Avoid "Spot" instances for critical tasks, as they can be interrupted.
Click "Edit Template" to configure permanent storage. Set the "pod volume" for storing models and outputs. The "container disc" is for temporary files and is deleted on pod restart.
Modify environment variables to enable or disable model and node installers in the startup script. Set parameters to "1" to install specific models or nodes.
Click "Deploy on Demand" to start the pod.
Network pods offer persistent storage across different GPUs.
Start the pod deployment process as usual.
Enable the "Network Volume" option and select a data center.
Give the volume a name and specify its size. Note that network volumes incur a monthly rental cost.
Create the network volume.
Deploy the pod, selecting any available GPU. Data saved to the network volume persists even after the pod is terminated.
Network volumes allow you to switch GPUs without losing your data.
Navigate to the desired folder (e.g., ComfyUI/models/Loras) and open a terminal.
Use the wget command followed by the download link from Hugging Face. Remove download=true from the end of the link.
wget <Hugging Face download link>
Rename the downloaded file if needed by right-clicking and selecting "Rename".
Navigate to the custom_nodes folder and open a terminal.
Use git clone to clone the custom node repository.
Use pip install -r requirements.txt to install any required packages.
pip install -r requirements.txt
Install packages like Sage Attention or Triton using pip install <package_name>.
Avoid using the ComfyUI manager to install pip packages, as it may result in security errors. Use Jupyter instead.
Watch the original video for visual context
This guide was automatically generated using AI (Google Gemini 2.0 Flash via OpenRouter) based strictly on the video transcript. All information comes directly from the video content. For visual demonstrations and additional context, watch the original video.
Generated on 10/16/2025 • Original video: Watch on YouTube