Reproduce “Unveiling the Secret of AdaLN-Zero in Diffusion Transformer” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jie Zhu, Mingyu Ding, Boqiang Duan, Leye Wang, Jingdong Wang

The authors examine how conditioning initialization affects training in DiT image-generation models and introduce adaLN-Gaussian, which initializes conditioning weights from a zero-mean Gaussian. Their comparisons suggest initial weight location matters more than update order. On ImageNet1K 256 × 256, it reaches FID 2.27 in 3,800K steps, matching a borrowed adaLN-Zero result at 7,000K steps; this is a step-count, not wall-clock, comparison.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.