Before Imagen and DALL-E 2 made diffusion the default for generative imagery, this Google Brain paper laid out the cascaded diffusion recipe that a lot of later text-to-image systems quietly borrowed: start with a small low-resolution diffusion model, then chain super-resolution diffusion models on top to upsample. The kicker is a technique called conditioning augmentation, which prevents compounding error between stages and turns out to be the difference between blurry mush and 4.88 FID at 256x256.

This long review of Ho et al. digs into the cascade design, the noise-conditioning trick that makes it robust, and the ImageNet benchmark results that beat BigGAN-deep and VQ-VAE-2 head-to-head. Speech folks should care: cascaded generation is now migrating into TTS and neural audio codecs, where multi-stage diffusion is being used to synthesize increasingly natural waveforms. If you want the foundations before touching modern audio diffusion work, the deep-dive walks through every component.