Custom Builds: run your own custom nodes and models on Comfy Cloud

Last updated: November 18, 2026

On Comfy Cloud, you can run Workflows that use your own Custom Nodes, models, and LoRAs, such as a private Node pack or a LoRA you trained. If a Workflow needs Custom Nodes that Comfy Cloud doesn't include, you package them in a custom Build and run it from the editor. For step-by-step setup, 👉 see Cloud Custom Builds in the Docs.

Q. What is a custom Build?

A custom Build packages the custom nodes (public or private), models, LoRAs, and Python dependencies your workflow needs. You deploy it to your own GPU workers on Comfy Cloud, then pick it under Run on in the Comfy editor. Your workflow then runs on your custom deployment instead of the default Comfy Cloud one.

Custom Builds use the same Builds and deployments as Comfy API, so one Build can serve both the editor and your own code.


Q. My workflow says custom nodes or models are missing. What are my options?

It depends on what's missing:

  • Only your own LoRA or model and every node is already on Cloud

    • Upload your model to the Model Library. No Build needed.

  • A custom node not available in Cloud

    • Create a Custom Build.

  • A node you'd like Comfy to support for everyone


Q. Can I use private or unpublished custom nodes?

Yes. A Build can include private custom nodes that aren't in the Comfy Registry. Only your Build uses them.


Q. Should I include models in the Build, or let them download?

  • Include in the Build: models you use on every run, especially large base models. They're ready when a worker starts.

  • Leave out: models you use occasionally. They download the first time a job needs them.

Important: Models in the Build add to storage costs for as long as its deployment exists, even when it's stopped. Models you leave out add download time to the first run on each new worker.


Q. Why is "Build with defaults" unavailable?

It's usually unavailable when Comfy can't match a node or model in your workflow to a known source. Click Customize build and add the missing node or model yourself.


Q. My Build failed to publish. What should I do?

  1. Open Logs on the failed Build and look for the first error.

  2. Fix the cause. Common causes are a Custom Node with dependency conflicts, a pinned package that doesn't install, or a model link that can't be reached.

  3. Edit the Build and click Publish again. Each attempt creates a new version.

Tip: Click Debug with your agent on platform.comfy.org to get a prompt you can paste into a coding agent to help fix the error.

Building is free, so failed attempts don't cost credits.


Q. How do I choose a GPU?

Pick the cheapest GPU with enough VRAM for your workflow.

  • If jobs fail with out-of-memory errors, redeploy on a GPU with more VRAM.

  • A larger GPU costs more per hour but may finish jobs faster.

Review the estimated cost before you click Create deployment.


Q. What do "Always-warm workers" and "Keep worker up after a job" mean?

  • Always-warm workers: With 0, you pay for GPU time only while a worker is up. That includes the Keep worker up after a job window after each job.

  • Keep worker up after a job: how long a worker waits for your next job before shutting down. The default of 3 minutes keeps back-to-back runs fast while you iterate.

For editor use, the defaults are usually best. 👉 For full billing details, see Comfy API: deploy your ComfyUI workflow as an API.


Q. Why is my first run slow?

When no worker is running, Comfy starts one, loads your Build, and downloads any models that aren't in it. Later runs are faster while the worker stays up.

Tip: To speed up first runs, include your main models in the Build.


Q. I can't see my deployment under "Run on".

Verify that:

  • The deployment status is Ready, not starting or stopped.

  • You selected Ready for the Cloud editor when you created the deployment.

  • You're signed in to the same account or workspace that owns the deployment.


Q. Does running on my deployment use my normal Cloud credits?

Yes. Runs on your deployment use your Comfy Credits, but they're billed for the deployment's GPU time and model storage instead of the usual Comfy Cloud usage. If you run out of credits, the deployment stops automatically.

Tip: Switching Run on back to Comfy Cloud returns to normal Cloud billing.


Q. Do Comfy Cloud concurrency and runtime limits apply to my deployment?

No. Your deployment has its own limits, separate from standard Comfy Cloud. Runs on your custom Build deployment does not count toward your Comfy Cloud concurrency limit.

Plan

Jobs at a time per deployment

Standard

2

Creator

2

Pro

10

Team

20

Tip: Jobs on a deployment can also run longer than the 30-minute and 1-hour caps on standard Comfy Cloud.


Q. Does switching "Run on" affect my other workflows?

Run on applies to the current editor session, not to a single workflow. If you have several Workflows open in the editor, they all run on the same deployment.


Q. My uploaded LoRA doesn't appear or won't import.

  • Use a .safetensors file of 2 GB or less. Put larger models in a Build instead.

  • Choose the matching model type when uploading (for example, LoRA).

  • Use a unique file name so it doesn't clash with an existing model.

  • Check your plan: uploads need Pro, Creator, Team, or Enterprise.

  • Make sure the LoRA matches your base model (for example, an SDXL LoRA with an SDXL checkpoint).

Note: Deleting a model from your Model Library is permanent, and jobs that still use it will fail.


Q. Can I use the same Build in the editor and from my own code?

Yes. Deploy the Build and call the deployment from code with the Python or TypeScript SDK or the HTTP API. 👉 See the Comfy API Quickstart.


Q. Can my teammates use my Build or deployment?

On Team and Enterprise plans, teammates in the same workspace can see your Builds. 👉 See Team workspaces: create, invite, and manage


Q. Where can I get help?

Related articles