Reproduce “G-STAR: Group-Scaled Token Capacity for Deeply Compressed Autoregressive Generation” 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
Xu Huang, Ye Huang, Xiaotong Li, Zhengrong Yue, Jialv Zou, Minhao Jing, Haodong Li, Wenhao Yan and 3 more
G-STAR shortens text-to-image generation by packing more discrete codes into each spatial step, preserving a fixed nominal information budget as spatial compression rises. Its semi-autoregressive head computes code representations in parallel, then predicts codes through a causal path. In a 512 × 512 ablation, it takes 3.41 seconds, faster than causal decoding and slower than parallel decoding, with near-causal quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.