Introduced on day 46
Diffusion model
A model trained to reverse a gradual noising of the data; generation starts from noise and denoises over many steps.
In detail
The forward process adds a little noise at each of many steps; a network learns to predict and remove the noise. Sampling runs the learned reverse process, which needs many network evaluations. Latent diffusion applies the process in the compressed space of an autoencoder to reduce cost, and guidance mixes conditional and unconditional estimates to trade variety for fidelity.