Reproduce “From Pixels, Without Pre-training: Joint Generative and Self-Supervised Representation Learning in One Model” 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
Vicente Balmaseda, Ching-Long Lin, Tianbao Yang
SCION learns image representations and generates images in one self-supervised, pixel-space model, without labels or pretrained components. A shared encoder uses clean augmented views for contrastive learning and noisy images for representation-conditioned generation; a learned prior supplies the representation at inference. On ImageNet 256 by 256, SCION reaches 8.92 FID versus RCG’s 14.27 in a matched 400-epoch comparison. This result is limited to that setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.