Skip to main content
Technology & EngineeringImage Generation Services32 lines

RunComfy ComfyUI Workflow Execution

Package a ComfyUI graph as a RunComfy workflow deployment and validate input overrides, asynchronous job state and output files.

Quick Summary9 lines
Confirm whether the task uses a ComfyUI graph deployment or a hosted model. RunComfy documents a serverless API for deployments and a model API for hosted model calls. Select the deployment workflow when the graph, node dependencies and named overrides are part of the artifact. Confirm project, billing scope and data handling before executing a remote request.

## Key Points

- [RunComfy API product selection](https://docs.runcomfy.com/)
- [RunComfy Serverless ComfyUI introduction](https://docs.runcomfy.com/serverless/introduction)
- [RunComfy workflow files and overrides](https://docs.runcomfy.com/serverless/workflow-files)
skilldb get image-generation-services-skills/runcomfyFull skill: 32 lines
Paste into your CLAUDE.md or agent config

RunComfy ComfyUI Workflow Execution

Confirm whether the task uses a ComfyUI graph deployment or a hosted model. RunComfy documents a serverless API for deployments and a model API for hosted model calls. Select the deployment workflow when the graph, node dependencies and named overrides are part of the artifact. Confirm project, billing scope and data handling before executing a remote request.

Package a graph

Start from a graph known to load in the intended ComfyUI environment. Record node and model versions, graph checksum, approved inputs and expected output type. Review custom nodes and source files before execution, especially if a graph came from someone else. Separate secrets from graph JSON and test payloads. Keep reference assets in approved locations and check the remote worker can read them.

Use the workflow format accepted by the target RunComfy operation. Identify the node IDs and input names for the values callers may override. Build a bounded override schema for prompts, dimensions, seed, input files and counts as appropriate. For one pilot, allow a single node input to change and verify the expected graph selection. Treat unexpected node IDs and file URLs as preflight errors.

Submit and track one pilot

Use the documented authentication and target deployment ID. Keep a local receipt with graph version, sanitized input summary and request identifier. Submit once and poll the documented status endpoint with bounded time and backoff. Fetch results only after a terminal success. If a response is lost after submission, inspect the known request or provider activity before starting another billed job. Report uncertain state explicitly.

Validate each returned artifact against the requested type, dimensions and expected content. Store retrieved outputs in a new job directory. Avoid logging signed delivery links or credentials. A queued response and a successful output download are different checkpoints.

Test failures before using a reusable endpoint

Try an unknown node override, inaccessible asset and a timeout in a mock or approved staging workflow. Confirm the resulting status is visible and failed requests cannot be marked complete. Verify deployment version and output association after an update. Keep rollback or previous-version routing documented when the workflow serves downstream callers.

Sources

Install this skill directly: skilldb add image-generation-services-skills

Get CLI access →