Stable Diffusion Training + Inference

Why it matters

SD family dominates open image generation. Understanding shapes vision generation.

The architecture

Text encoder: CLIP or T5.

UNet: denoises latents.

VAE: latent to pixel.

How it works end to end

Training: predict noise added to latent.

Inference: 20-50 denoising steps.

Schedulers: DDIM, DPM++.

ControlNet: conditioning.