DiffusionBlocks: Training Neural Networks One Block at a Time

I do not understand.

how is this different from building smaller transformer layers, and each layer just denoises less?