Radial gate objective cuts diffusion transformer FID from 16.24 to 11.83
Junha Park · hf · 2026-11-02
- Key finding: Training acceleration in diffusion transformers stems from reorganization of AdaLN-Zero residual gates; broader gate participation (measured by a new nPR metric) strongly correlates with generation quality (Pearson r = -0.96 on SIT-XL/2 between nPR and FID).
- Method: A simple radial objective promotes broader per-sample gate participation without auxiliary teachers.
- Results: On SIT-B/2·400K, FID drops from 16.24 (vanilla) to 13.19 (dispersive) to 11.83 (radial), with nPR rising from 0.434 to 0.642.
More from Research
- General Reasoning explains how to train agent swarms with RL and latency pressure — rosstaylor90 · 2026-11-02
- Mathematician Tony Feng: current-paradigm LLMs lack the diversity of thought research needs — littmath · 2026-11-02
- Measure Theory vs Probability: 'You're Confusing the Finger With the Moon' — LucaAmb · 2026-11-02
- DStill Distills Multi-Step Denoising Into One Pass, 6.45x Speedup for Diffusion Speculative Decoding — LucaAmb · 2026-11-02
- REA: async brain-cerebellum runtime for robots where policies never wait on the LLM — Pengfei Zhou · 2026-11-02
- Microsoft's MemSteer turns unlabeled user histories into agent memory supervision — MicrosoftResearch · 2026-11-02