Reproduce “REPA Works Until It Doesn't: Early-Stopped, Holistic Alignment Supercharges Diffusion Training” 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
Ziqiao Wang, Wangbo Zhao, Yuhao Zhou, Zekai Li, Zhiyuan Liang, Mingjia Shi, Xuanlei Zhao, Pengfei Zhou and 4 more
HASTE trains image diffusion transformers by combining pretrained vision-model feature and attention-map guidance early, then removes both auxiliary losses and continues with denoising alone. On ImageNet 256×256, SiT-XL/2 with HASTE reaches FID 8.39 after 50 epochs, versus FID 8.61 after 1,400 epochs for vanilla SiT-XL/2—about 28 times fewer optimization steps. This is a step-count comparison, not a wall-clock speedup.
Key insights
- External guidance from pre-trained models like DINOv2 accelerates early diffusion model training but becomes detrimental later when the model develops its own generative capabilities, due to conflicting objectives and capacity mismatch.
- Analyzing the cosine similarity between denoising loss gradients and external guidance loss gradients reveals when external guidance shifts from helpful (positively correlated) to harmful (negatively correlated), indicating an optimal termination point.
- Attention patterns and feature representations from a teacher model can be effectively transferred to a student diffusion model, providing valuable relational priors that accelerate learning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.