🪢Diffusion models

Diffusion models are a class of generative models that convert Gaussian noise into samples from a learned data distribution via an iterative denoising process. These models can be conditional on class labels, text, or low-resolution images. A diffusion model is trained on a denoising objective of the form:

Where are data-conditioning pairs, , and are functions ofthat influence sample quality. Intuitively, is trained to denoise intousing a squared error loss, weighted to emphasize certain values of. Samplings such as the ancestral sampler and DDIM start from pure noiseand iteratively generate points , where , that gradually decrease in noise content. These points are functions of the x-predictions .

Classifier guidance is a technique to improve sample quality while reducing diversity in conditional diffusion models using gradients from a pre-trained model during sampling. Classifier-free guidance is an alternative technique that avoids this pre-trained model by instead jointly training a single diffusion model on conditional and unconditional objectives via randomly droppingduring training (e.g. with 10% probability).

Sampling is performed using the adjustedprediction , where

Here, and are conditional and unconditional predictions, given by , andis the guidance weight. Settingdisables classifier-free guidance, while increasingstrengthens the effect of guidance. Imagen depends critically on classifier-free guidance for effective text conditioning.