Every step — the model, the sampler, the seed, each adapter and its weight, the upscale, the mask — is a node you place and connect, and nothing happens that is not on the canvas. That is tedious for one picture and indispensable for a set, because a graph is a thing you can save, diff, hand to somebody else and run again in six months to get the same output.
The artwork on this site was built this way, and the write-up of how is in the lab. The reason it was worth the setup is repeatability: a single room needed dozens of layers and props that had to look like they belonged together, and that is a pipeline problem rather than a prompting one.
Related terms
Stable Diffusion
A family of open-weight image models that generate a picture by repeatedly removing noise from a random start, guided by a text description.
LoRA (low-rank adaptation)
A way of fine-tuning that trains a small add-on layer instead of the whole model, producing a file of a few megabytes rather than a few gigabytes.
The bench this belongs to
Generative AIStable Diffusion and ComfyUI wired into the place where your content actually gets made, with a model fine-tuned so everything comes out looking like you.
