Reproduce “Elucidating the Preconditioning in Consistency Distillation” 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
Kaiwen Zheng, Guande He, Jianfei Chen, Fan Bao, Jun Zhu
Analytic-Precond changes how consistency trajectory models combine a noisy input with a denoiser prediction, aiming to improve learning of intermediate jumps along a diffusion teacher’s ODE trajectory. The authors derive the weights from that trajectory and minimize a bound on the consistency gap. In two-step CTM training, CIFAR-10 class-conditional curves show a 2.8× iteration speedup to comparable FID; final gains are small and nonuniform.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.