Diffusion Model

What Diffusion Model is

Diffusion replaced GANs as the dominant image generation approach because it trains stably and produces diverse, high-fidelity results.

How it works

Training adds noise to real samples in steps and teaches a network to predict and remove it. Generation starts from pure noise and denoises iteratively, guided by a text embedding through cross-attention. Latent diffusion runs this in a compressed space for speed, and distilled samplers cut the number of steps dramatically.

Why it matters

Nearly all current text-to-image and much text-to-video generation is diffusion-based, and the same machinery is spreading to audio and molecular design.

Common uses

• →Text-to-image generation
• →Inpainting and outpainting
• →Video generation
• →Audio and music synthesis