Reproduce “Learning Energy-Based Models by Cooperative Diffusion Recovery Likelihood” 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
Yaxuan Zhu, Jianwen Xie, Yingnian Wu, Ruiqi Gao
CDRL trains an initializer to propose lower-noise images for an energy-based model (EBM) to refine with Langevin updates. The initializer learns from refined proposals, absorbing some sampling transitions without replacing the EBM. On 32×32 CIFAR-10, CDRL reaches FID 4.31 versus DRL’s 9.58, using 15 rather than 30 MCMC steps per noise level; it does not outperform all generative models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.