15 upvotes
Residual Gates, Not Regularization, Speed Up Diffusion Transformers
What is No Channels Left Behind: Understanding Diffusion Acceleration through Residual Gates?
The paper shows that acceleration tricks for diffusion transformers (like REPA) work mainly because they change the AdaLN-Zero residual gate scaling, not because of extra supervision or spatial priors. By tracking the gate’s α value the authors explain the internal shift that makes training faster.
Who it is for
Anyone training diffusion models can tune the residual gate instead of adding complex regularizers to cut training time.