Reproduce “Just Image Transformers with Late-Stage Pixel Refinement” 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
Denis Lukovnikov, Andreas Müller, Asja Fischer
The paper extends Just Image Transformers (JiT) with a convolutional pathway that refines local detail late in denoising, while an overlapping linear head handles early stages. JiT predicts the clean image directly from noisy input. On class-conditioned ImageNet at 256×256, JiT-H/16 reaches gFID 1.78 versus 2.34 for a corrected JiT reproduction; this is not a best-in-class claim.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.