Study Reveals Residual Gating Channel Utilization Drives Diffusion Acceleration

A new arXiv paper shows that diffusion Transformer training acceleration stems from reorganized AdaLN-Zero residual gating. The proposed nPR metric correlates gate participation breadth with generation quality, and a radial objective reduces FID to 11.83.

2026-11-02 ~ 2026-11-03 · 2 related posts

• Radial gate objective cuts diffusion transformer FID from 16.24 to 11.83 — Junha Park · 2026-11-02
• Paper Links Diffusion Training Speedups to Residual Gate Channel Use, Improves FID Teacher-Free — serrjoa · 2026-11-03