GENERATIVE AI

Comfy Cloud Adds Custom Builds That Run Teams' Own ComfyUI Setups in the Browser

Open flight case labeled Comfy Cloud floating over sunset clouds, holding node graphs and wired to three laptops showing the same graph.

Comfy added custom Builds to Comfy Cloud, letting a team package its own ComfyUI nodes, models, LoRAs and private node packs into one setup that every artist opens in the browser, as the company said on X. Studios now keep one shared ComfyUI environment in the cloud instead of maintaining a separate setup on every artist's machine.

  • Private node packs are supported. According to Comfy Org, a Build bundles custom nodes from the Comfy Registry or GitHub, uploaded private or internal packs, models, LoRAs and Python dependencies.
  • Updates reach the whole team at once. Add a new LoRA to the Build and roll out an update, and every artist works from the same setup, including new artists on their first day.
  • GPU choice runs up to the 180GB B200. Teams pick the hardware for each deployment, from the RTX PRO 6000 in 48GB or 96GB versions up to H100, H200 and B200.
  • Concurrency is set by plan. Each deployment runs up to 2 jobs at a time on Standard and Creator, 10 on Pro and 20 on Team, and scales back down when jobs finish.
  • Developers call the same Build by API. Pipeline tools can run the workflows artists edit on the canvas through the Developer Platform API.

A custom Build packages a team's ComfyUI environment once, and the whole workspace opens it in the browser

According to Comfy's announcement, setup usually happens once, handled by whoever owns the team's ComfyUI pipeline. They start on the Developer Platform from a workflow or a Comfy Desktop snapshot, and the builder finds the custom nodes and models it needs and helps resolve dependency conflicts.

After choosing a GPU and deploying, anyone in the workspace can pick the Build in the Comfy Cloud editor and adjust prompts, parameters and connections on the familiar canvas. Teams can create or edit different workflows on one Build without rebuilding it, as long as the nodes and models those workflows need are already included.

The Comfy Cloud page says some node behaviors aren't supported yet, such as nodes that wait for a browser click during a run, and Comfy notes complex packs can take some troubleshooting. The company recommends testing each workflow on its deployed Build before relying on it.

People outside the workspace can't open a Build, even with its URL.

Comfy logo: the word Comfy in bold black stylized letters on a slanted neon-yellow badge over a backdrop of pink-tinted clouds and blue sky
The Comfy logo from the company's custom Builds announcement video. Image: X (formerly Twitter)

GPU time bills by the second, and an open canvas doesn't start a GPU

Custom Builds are included in every paid Comfy Cloud plan, starting with Standard. GPU time costs $1.75 to $11.23 an hour depending on the GPU, billed by the second while workers are running, and Comfy lists every rate and plan limit on its pricing page.

Opening or editing the canvas doesn't use GPU time. Charges do apply during a keep-alive window or while workers are kept warm, and when no jobs are running or queued, compute can sleep according to the deployment's idle settings, so the next run may take longer to start.

A team's models live on network storage billed monthly, whether or not anything is running. Plans also set how many custom Builds a team can keep.

For the remaining setup, Comfy points teams to its documentation, which it says covers packaging from the command line, rolling out updates and running workflows by API. The general Comfy Cloud docs explain running ComfyUI workflows online without a local install.

Standard and custom setups now sit a click apart in the same Comfy Cloud editor

Comfy Cloud, whose public beta we previously covered, now offers both the standard setup and a team's custom Build in one editor. Custom Builds cover the custom nodes, private packs and LoRAs that Comfy says workflows rely on beyond the default nodes.

Inviting teammates into a shared workspace requires the Team plan, which also carries the highest limit of 20 concurrent jobs per deployment.

KEEP READING