Diffusion model
Generative AI model trained to progressively remove noise from data. The inverse process — add noise to a clean image → model learns to recover — gives generation. Drives all modern image/video models.
Technical milestones: Score Matching (Hyvärinen, 2005), DDPM (Ho et al., 2020), Latent Diffusion (Stable Diffusion, 2022), Diffusion Transformer (DiT, 2023 — powers Sora). Distinct from GAN: more stable training, higher quality, slower inference. Sampling steps are typically 20–50 but can be reduced via distillation (Latent Consistency Models, SDXL Turbo).