ComfyUI model tutorial: LoRA, CLIP, VAE and ControlNet

This beginner chapter explains the model files behind ComfyUI image and video workflows. It starts with diffusion models, then describes how LoRAs modify a base model for a particular person, style or subject. The speaker also explains why storage and GPU memory matter when you run models locally, describing modern diffusion files as commonly 10 to 20 GB, with some reaching 40 GB or more. These are the speaker's general estimates, rather than requirements for every workflow.

The tutorial separates text encoding from image decoding. CLIP converts a prompt into numerical conditioning, while a VAE converts between latent data and visible pixels. The speaker recommends checking the diffusion model's documentation for compatible files and explains that mismatched components can cause errors or poor output. Folder guidance includes models/clip and the VAE folder under models.

ControlNet gets a longer practical explanation: a reference image passes through a preprocessor to supply pose, depth or edge information that guides composition. The chapter identifies models/ControlNet as the destination for those files. It closes with a brief discussion of Stable Diffusion 1.5 and SDXL, then notes that newer models differ in licensing, LoRA support and intended use. The emphasis is on understanding component roles rather than choosing a single best model.