ML//diffusion model//consistency model

- Distillation: train a model to map any noise level directly to the clean image in one step.

Distillation: train a model to map any noise level directly to the clean image in one step.

Standard diffusion needs 20–50 denoising steps; consistency models need 1–4, orders of magnitude faster.

LCMs (Latent Consistency Models) enabled the "quick draw" viral moment (Dec 2023), real-time generation.

Updated with sCMs and DMDs. Core tradeoff: sacrifice some quality for massive speed gains.