ComfyUI

ComfyUI is a node editor for image models. You load a checkpoint, wire a prompt into a sampler, wire the sampler into a decoder and press run. That sounds like extra work compared with a text box, and for the first week it is. After that the graph becomes the point: every step is visible, every step can be swapped, and the whole thing saves as a JSON file you can send to someone else.

How I use it

I run it on my Mac for the jobs where volume matters more than polish. When I need forty variations of a product scene to pick three from, paying an API per image adds up fast and the quality difference is small. I keep a handful of saved workflows: one for plain generation, one for inpainting a region of an existing image, one for upscaling. Most of my sessions are the inpainting one, fixing a hand, a label or a background corner on an image that is otherwise fine.

Because workflows are files, I can run them from a short script instead of clicking through the interface. That turned it from a toy into a tool.

What surprised me is how quickly new techniques appear here. A paper lands, and a custom node for it exists a few days later, long before any hosted product offers it. The flip side is that those nodes are maintained by one person each, and an update to the core can break a third of them in one go. I keep a known-good install and update on purpose, never by accident.

What annoys me: speed on Apple silicon is fine for stills and poor for video models. Some actions have no undo. Error messages are Python tracebacks, which is fine for me and a wall for most people.

Who it is for

Developers and technical designers who generate a lot of images and want control over the steps. If you need one hero image a week, a hosted tool with a text box will make you happier. If you want to understand what the model is doing and stop paying per image, start here and give it a weekend.