Reading the graph

Everything in ComfyUIComfyUIThe dominant node-based interface for running diffusion models locally. Started by 'comfyanonymous' in January 2023, now stewarded by Comfy Org. is a graph of nodes joined by typed wires. A nodeNodeA single block in a visual pipeline like ComfyUI that does one operation and wires to others to form a workflow. is one operation with inputs on its left and outputs on its right; a wire carries one kind of data from an output to a matching input, and the colors are the type system. Six types come up constantly: MODEL (the diffusion network itself), CLIPCLIPA text encoder developed by OpenAI in 2021 that learns to align text and images in a shared embedding space. Foundation of most text-to-image models from 2022 onward. (the text encoderText encoderA neural network that converts text prompts into numerical embeddings that the diffusion model can use. Modern models use combinations including CLIP and T5.), VAEVAE (Variational Autoencoder)A neural network that learns to compress images into a small numerical representation (a latent) and decompress them back. Used in latent diffusion models to make the generation process tractable. (the translator between pixels and latent spaceLatent spaceThe compressed numeric representation an image is turned into before diffusion. Working here instead of on raw pixels is what makes modern image models fast enough to run.), CONDITIONINGConditioningAny extra input that steers generation beyond the prompt, such as a depth map, pose, edge map, or reference image. (an encoded promptPromptThe text description you provide to a model to specify what you want it to generate.), LATENT (an image while it is still in the model's compressed working space), and IMAGE (actual pixels). A wire only connects an output to an input of the same type, which makes the graph self-documenting: if a connection is legal, it is because the data genuinely fits there.

You run a graph by queuing it. ComfyUI does not recompute the whole thing each time; it caches and re-executes only the nodes downstream of whatever you changed, which is why iterating on the last step of a long pipeline stays fast. And because a finished image carries its own recipe, dragging a PNG that ComfyUI generated back onto the canvas rebuilds the entire workflow that made it, seeds and all. Workflows are just JSON, so they travel as files or ride along inside the images themselves. That single fact, that an output is also its own reproducible source, is one of the quiet reasons the tool won.