Cascaded diffusion chains models at increasing resolutions, using noise-conditioning augmentation between stages

To generate very high-resolution images, cascaded diffusion (Ho et al. 2021) chains multiple diffusion models: a base model generates a small image, then successive super-resolution models upsample it. The critical technique between stages is noise-conditioning augmentation: the conditioning image fed to each super-resolution model is corrupted with Gaussian noise or blur during training. This stops the super-resolution model from relying on high-frequency artifacts produced by earlier stages, reducing compounding error through the cascade. The augmentation is applied only during training, not at inference (though the augmentation level is signalled to the model). The most effective corruption is Gaussian noise at low resolution and Gaussian blur at high resolution.

Examples

Imagen (Saharia et al. 2022): a 64x64 base model, then 64->256 and 256->1024 super-resolution diffusion models. Each stage carries its own text conditioning.

Assessment

What is noise-conditioning augmentation in cascaded diffusion? Explain what would go wrong without it and why different corruption types are used at different scales.