Key DDPM follow-up works include improved variance learning, cascaded generation, classifier guidance, and classifier-free guidance

The original DDPM was the seed for a rapid sequence of improvements. Nichol et al. (2021) added learned variance alongside the mean, improving sample quality. Ho et al. (2021) introduced cascaded diffusion — a pipeline of multiple diffusion models at increasing resolutions. Dhariwal et al. (2021) showed diffusion models beat GANs via architecture improvements and classifier guidance (using a separately trained classifier gradient at inference). Ho & Salimans (2021) introduced classifier-free guidance, which avoids needing a classifier by jointly training conditional and unconditional models and blending predictions at inference. DALL-E 2 and Imagen then combined these with large language/CLIP models for text-to-image generation.

Examples

Classifier guidance: add a classifier’s log-probability gradient to the denoising step. Classifier-free: interpolate between conditional and unconditional model outputs at inference time.

Assessment

List four DDPM follow-up papers and describe the key contribution of each. Explain the difference between classifier guidance and classifier-free guidance.