I built an agent-first CLI that deploys a RunPod serverless ComfyUI endpoint and runs workflows from the terminal (plus a visual pipeline editor)

Developer Hearmeman98 has released two complementary open-source tools, ComfyGen and BlockFlow, designed to move AI image and video generation workflows from local ComfyUI instances to scalable, serverless GPU infrastructure on RunPod. The core tool, ComfyGen, is an agent-first command-line interface (CLI) that automates the deployment of a serverless ComfyUI endpoint. It features an interactive setup wizard, can download models and LoRAs directly to the GPU's network volume via serverless jobs, and executes workflows by uploading inputs to S3 and polling the endpoint. Crucially, it outputs structured JSON, making it ideal for automation by AI coding agents.

BlockFlow is a locally-run, browser-based visual pipeline editor that allows users to chain generation steps—like LLM prompt writing, ComfyUI execution, and upscaling—into complex, branching workflows. While the tools can be used together, with BlockFlow orchestrating pipelines that ComfyGen executes, the CLI is also designed for standalone scripting. This serverless approach means GPUs spin up only when a workflow runs, enabling users to scale large batches across multiple GPUs and pay only for execution time, eliminating the need to manage and pay for idle GPU pods.