Reproduce “FasterDiT: Towards Faster Diffusion Transformers Training without Architecture Modification” 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
Jingfeng Yao, Wang Cheng, Wenyu Liu, Xinggang Wang
Huazhong University of Science and Technology researchers propose FasterDiT, a training strategy that significantly accelerates Diffusion Transformer convergence without architectural changes. It achieves a 7-fold speedup, enabling DiT-XL/2 to reach a competitive FID of 2.30 on ImageNet 256x256 in 1 million iterations, compared to the original DiT's 7 million, by optimizing SNR distribution and introducing a velocity direction loss.
Key insights
- The absolute Signal-to-Noise Ratio (SNR) during training is influenced by the input data's signal strength (standard deviation), which significantly impacts training outcomes and model robustness.
- Training strategies with broader SNR PDF distributions tend to exhibit greater stability and robustness across varying data signal intensities.
- An 'optimal SNR range' exists where the mean of the SNR PDF correlates with improved FID performance, providing a unified target for training optimization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.