ComfyUI for Image Generation

This chapter examines the foundational mechanics of ComfyUI workflows, focusing on how base models, prompts, and sampling parameters interact to generate images. It begins with an explanation of checkpoints—pretrained diffusion models containing the learned parameters of networks such as Stable Diffusion—and their critical components: UNet, CLIP text encoders, and VAEs. The chapter details how prompts are encoded into vector embeddings that guide the generative process, with negative prompts used to suppress unwanted features. Latent initialization and seed control are introduced as techniques for ensuring reproducibility and controlled variation. Particular emphasis is given to the KSampler node, which orchestrates iterative denoising according to parameters such as sampling steps, CFG scale, and denoising strength. Examples illustrate how these variables shape architectural outputs, from schematic ambiguity to spatial precision. The chapter concludes by situating sampling as analogous to design iteration: early low-step renders provide conceptual sketches, while high-step, tightly constrained generations yield detailed, contextually coherent images. For architects, mastering these mechanisms transforms ComfyUI from a technical tool into a medium for intentional and repeatable design exploration.