Reproduce “BudgetPix: Compute-Adaptive Tokenization for Pixel-Space Image Diffusion” 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
Ozgur Kara, Yujia Chen, Daniel Watson, David Forsyth, James Matthew Rehg, Wen-Sheng Chu, Du Tran
BudgetPix adapts pretrained pixel-space diffusion models to variable token budgets. It uses an entropy-guided quadtree, multi-scale encoder and decoder, and training and sampling that allocate larger tokens to simple regions and finer tokens to detail. On MiniT2I-L/16, it scores 0.874 GenEval at 25% tokens, versus 0.882 full-budget; measured speed-up is 3.59×, not fourfold. PixelDiT-T2I shows 1.95× due to its full-resolution pixel stage.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.