Reproduce “Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots” 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
Guangting Zheng, Yehao Li, Yingwei Pan, Jiajun Deng, Ting Yao, Yanyong Zhang, Tao Mei
Hi-MAR introduces Hierarchical Masked Autoregressive Models for visual generation, addressing limitations in capturing global context and information loss by employing a two-phase process with continuous-valued tokens. The model achieves state-of-the-art FID scores on ImageNet (1.52) and MS-COCO (4.77), outperforming leading diffusion and autoregressive models while requiring 54% fewer resources than its MAR baseline.
Key insights
- Hierarchical generation guided by predicted low-resolution tokens effectively addresses the global context problem and ensures consistency between training and inference.
- Modeling inter-token dependencies with a Diffusion Transformer head significantly enhances masked token prediction and image coherence.
- Incorporating scale-aware mechanisms allows a shared Transformer backbone to efficiently process information across different resolutions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.