Choose a compatible GPU and ComfyUI template, attach storage deliberately, open the running interface and test one small workflow. Check compute and storage charges separately and copy important outputs to durable storage.
Watch on YouTube · Video published 2025-11-17 · Companion reviewed 2026-09-23
What this video covers
This is a RunPod setup lesson published in November 2025. It demonstrates running the ComfyUI interface on a cloud GPU. It does not demonstrate the Comfy Cloud API or a finished API business. Use the separate Build & Earn guide for the service-design and API discussion.
Plan the session before starting a GPU
- Choose the exact workflow and model first. Check its memory, disk and dependency requirements; a cheap hourly GPU is not automatically suitable.
- Estimate model downloads and output storage. The video’s small example storage allocation should not be treated as sufficient for modern video models.
- Decide where files must survive. RunPod container disk, pod volume disk and network volume have different lifecycles. Confirm the current behavior in the provider documentation.
- Read the displayed compute and storage rates. Include model download time, idle time, retries and retained storage in the session budget.
Follow the setup sequence
- Choose a region and storage arrangement compatible with your intended GPU. Record the volume and pod names so you can identify them later.
- Select a ComfyUI template. The video uses DossAI_ComfyUI_Manager; templates and included node versions can change. Review the template details before deploying.
- Start the pod and wait for the provided service connection. Open ComfyUI through the provider’s connection interface. Keep authentication and private inputs out of public workflow exports.
- Load one basic graph. Install its required models and nodes, select the files in the loaders and run a small image before trying a large video workload.
- Download a test output and save the workflow, model versions and input list. Confirm that your intended persistent storage actually contains the files you want to retain.
- When finished, stop the billed compute resource as appropriate. Check the dashboard for continuing storage charges and verify your files before terminating or deleting anything.
Two corrections to the older video
Storage: stopping compute is not the same as deleting a network volume. Persistent storage can keep billing while the GPU is stopped. Conversely, some container data is ephemeral. Treat the provider’s current lifecycle rules as authoritative, and test persistence with a disposable file before relying on it.
Seeds: zero is a valid seed value. Use the workflow’s randomize/increment control when you actually want a changing seed. Entering zero alone does not universally mean “random.”
Cloud troubleshooting
- Interface does not open: check the pod’s status, template startup output and exposed service link. Do not assume the model itself is the problem.
- Missing nodes or models: compare the actual environment with the workflow’s dependency list. A template is a starting environment, not a guarantee of zero installation errors.
- Out of memory: lower resolution, duration or batch size, or select hardware appropriate for the graph.
- Unexpected bill: check both running compute and storage. A stopped GPU does not prove all charges have stopped.
Keep a useful session record
Record which GPU you used, the model filenames, generation duration, successful outputs and failures. That record helps compare cloud and local costs fairly. For client work, also record your quality-review time and delivery overhead. A fast sample generation alone is not a complete service cost. Old promotional credit amounts from the video are not offered here; check the provider’s current terms directly.
Sources and downloads
- RunPod storage types and persistence
- RunPod current services and pricing
- Published cloud setup video
- From workflows to services