Paper ties diffusion training speed to residual-gate layout
Original title: No Channels Left Behind: Understanding Diffusion Acceleration through Residual Gates
A teacher-free radial objective on ImageNet 256 with SiT-B/2 improves FID over dispersive loss while widening gate participation.
The paper argues that diffusion-transformer acceleration methods such as REPA and dispersive regularization reorganize the AdaLN-Zero residual gate, not only the features those losses supervise. Accelerated models are more sensitive to perturbations of that gate, and broader channel participation, measured by a normalized participation ratio, tracks better generation quality. The authors add a teacher-free radial objective that regularizes feature norms without pairwise comparisons. On ImageNet 256 with SiT-B/2, they report better FID than dispersive loss and higher mean gate participation across timesteps and classes.
- The gate studied is the AdaLN-Zero residual scale that admits conditioned updates.
- nPR measures how broadly channels participate in the gates.
- The radial objective needs no teacher and no pairwise feature comparisons.
- Experiments use ImageNet at 256 resolution and the SiT-B/2 backbone.
- The claim is that gate reorganization is one measurable cause of faster training.
From the original article
1 min read at huggingface.co
Training acceleration methods for diffusion transformers, such as REPA and dispersive regularization, are commonly explained by additional supervision, spatial-structure priors, or improved feature geometry.
By Junha Park, Jaemyung Yu, Jaehui Hwang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.