Diffusion models overview

Diffusion models learn to generate images (and other media) by practicing how to remove noise step by step. At generation time they start from noise and denoise toward a picture—often guided by a text prompt.

Diffusion underpins many popular image tools. Knowing the sketch helps you debug weird outputs and understand why sampling steps/cost matter.

Training: take real images, add noise, teach a network to predict/clean the noise. Generation: start from random noise, repeatedly clean a little, optionally steered by text embeddings.

A fogged window slowly wiped clear—except the “clear picture” is invented to match the prompt’s statistics.

Generate the same prompt twice; note what stays stable vs random. That residual randomness is sampling, not a unique “correct” image.

⚠️ Myth: More steps always mean a better image forever.

✓ Reality: Diminishing returns; some distillations use few steps.

⚠️ Myth: Photoreal equals authentic photograph.

✓ Reality: See Course 29 deepfake literacy.

• Ho et al. diffusion papers (cite the specific one you teach)
• Course 11 generative-media-overview; Course 08 multimodality
• Model cards for the tool you demo