No Channels Left Behind: Understanding Diffusion Acceleration through Residual Gates

No Channels Left Behind. Understanding Diffusion Acceleration through Residual Gates. By Junha Park and four others, from Yonsei University and NAVER AI Lab. Published on the twenty-ninth of October.

Methods that speed up the training of diffusion transformers, such as REPA and dispersive regularization, are usually explained by what they add: extra supervision, priors about spatial structure, or better feature geometry. This work revisits that account. Under these methods, a single internal component is substantially reorganised: the AdaLN-Zero residual gate, alpha, which directly controls how conditioned updates enter the residual stream. The hypothesis that follows is that acceleration comes, at least in part, from reshaping this gating pathway, and not solely from the feature structures the auxiliary objectives explicitly impose. Put plainly, the usual story looks at the features these methods shape, and this one looks at the valve that decides how much of each update gets through.

Several observations support it. Accelerated models are markedly more sensitive when alpha is perturbed, and the geometry of their gates reflects semantic relationships among classes. To describe how the gates are organised across channels, the authors introduce the normalized participation ratio, or nPR. Across REPA teachers, projector architectures and radial-objective targets, broader gate participation goes together with better generation quality. The pattern holds across those three kinds of variation rather than in a single configuration, which is what lets it serve as a guide.

That relationship then guides a new objective. It is a simple radial objective that needs no teacher, and it promotes broader gate participation through feature-norm regularization, without any pairwise comparison of features. On ImageNet at two hundred and fifty-six by two hundred and fifty-six, with SiT-B/2, it improves FID over dispersive loss, while raising mean gate participation over timesteps and classes. Both movements happen together: quality goes up and participation broadens, which the authors take as further support for the link between how the gates are organised and how fast training proceeds.

The claim is stated with care. Acceleration is said to arise from the gates at least in part, so the auxiliary objectives' own feature structure is not ruled out. The link between participation and quality is an accompaniment observed across settings rather than a proof of cause, and the new objective is reported on one image resolution with one model. What is established is that residual-gate reorganisation is a measurable axis of diffusion training acceleration, and that it suggests a simple direction for going further.

The title makes the point in four words: no channels left behind. Under the methods studied, faster training goes with gates that spread their participation across more channels, and a regulariser that pushes for exactly that, with no teacher at all, already beats dispersive loss on FID.