the Diffusion Transformer replaced the U-Net denoiser because it scales predictably with compute; the U-Net survives in small and on-device models where quadratic attention does not fit
Learn more →
Why study this historical topic?
the Diffusion Transformer replaced the U-Net denoiser because it scales predictably with compute; the U-Net survives in small and on-device models where quadratic attention does not fit
The U-Net Used as the Denoiser in a Diffusion Model
a segmentation network repurposed to predict noise was the default denoiser for years, and knowing why it was chosen is what makes its replacement legible
Diffusion Model Forward and Reverse Processes using UNets in Machine Learning
Diffusion models represent a generative framework grounded in stochastic processes where data distribution is reversed by learning to predict and remove noise added via a fixed forward Markov chain. …