Reproduce “Just Align $\bm{x}$: Aligning Predictions, Not Representations” 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
Yuyao Zhang, Yuwei Hu, Ziyang Mai, Yu-Wing Tai
JAx aligns clean-image predictions across noise levels to train pixel-space JiT generators, which predict a clean image from a noisy input. An EMA teacher supplies predictions from cleaner observations coupled to student inputs, alongside direct image supervision. On ImageNet 256×256, JAx lowers FID at all three scales and reaches baseline 600-epoch FID after 400 epochs, not establishing lower elapsed time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.