ComfyUI

Install the desktop app (the easiest entry point) or set up manually with Python 3.12+ and PyTorch → build a node graph: a model loader, a prompt or image input, generation nodes, an output → run it on a small test case → save the workflow once it produces something usable → reuse or adapt it for the next similar task.

ComfyUI itself is free, open-source software; you pay for the machine, specifically a GPU capable of running the models you choose, since CPU-only generation is slow enough to be impractical for most real work. If you don't have that hardware, Comfy Cloud is the official paid option, billed for cloud compute instead of a local GPU. Optional paid "API nodes" let a local workflow call closed-source hosted models when you want one without leaving the graph, billed separately per call.

ComfyUI's own code is GPL-3.0; individual model weights you load into it carry their own separate licences and usage terms. Running non-trivial models needs real GPU memory; check a model's documented requirements before assuming your hardware can run it.

Build a simple three-node image generation workflow: a model loader, a text prompt node with a made-up description, and a save-image output. Run it once, then change only the prompt and run it again to see how much of the pipeline you can reuse unchanged.

Done when you've generated two different images from the same base workflow by changing only the prompt node, and you understand which parts of the graph you'd need to change for a genuinely different task.