Reproduce “Unlocking Any-Order Generation in Pretrained Autoregressive Image Models” 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
Rishav Pramanik, Marco Pedersoli, Zhaozheng Yin
The authors adapt raster-scan image generators for image completion in arbitrary token orders, adding a target-location signal and gradually shifting training from raster order to random permutations. Across four ImageNet models, adaptation uses 13–33% additional training compute; LlamaGen-XL’s class-conditional FID rises from 2.63 to 3.91. Inpainting FID is 6.21 versus MaskGIT Halton’s 7.82, but evidence is limited to class-conditional ImageNet tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.