Reproduce “DINO-Tok: Adapting DINO for Visual Tokenizers” 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
Mingkai Jia, Mingxiao Li, Zhijian Shu, Anlin Zheng, Liaoyuan Fan, Jiaxin Guo, Tianxing Shi, Dongyue Lu and 7 more
DINO-Tok adapts frozen DINO features for continuous and discrete image tokenization. It combines deep semantic features with shallow detail features, and its discrete version quantizes PCA-selected semantic dimensions alongside texture codes. On ImageNet-1k at 256×256, its 1.4B-parameter generator reports gFID 2.44 and IS 254.2, versus 3.08 and 253.6 for LlamaGen-XXL; these results cover this setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.