[Short Review] Cascaded Diffusion Models for High Fidelity Image Generation
Cascaded diffusion is the two-stage-plus recipe underneath a lot of modern generative work: generate a small image with one diffusion model, then hand it to super-resolution diffusion models that add high-frequency detai
Cascaded diffusion is the two-stage-plus recipe underneath a lot of modern generative work: generate a small image with one diffusion model, then hand it to super-resolution diffusion models that add high-frequency detail. The Google Brain team showed the whole pipeline hitting 4.88 FID at 256x256 on ImageNet, no auxiliary classifier boosting required, and beating BigGAN-deep in the process.
This short review of Ho et al. covers the essentials fast, the cascade structure, why the conditioning augmentation trick is the secret sauce that keeps errors from compounding across stages, and how the FID numbers stack up. It's worth 5 minutes for anyone whose TTS or neural vocoder roadmap includes multi-stage diffusion; the same design pattern is showing up all over audio generation now. Give it a spin as a quick primer before jumping into the full paper.