Reproduce “Designing a Conditional Prior Distribution for Flow-Based Generative Models” 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
Noam Issachar, Mohammad Salama, Raanan Fattal, Sagie Benaim
Researchers at The Hebrew University of Jerusalem introduced a Conditional Prior Distribution (CPD) for flow-based generative models, replacing generic noise priors with condition-specific Gaussian distributions. This modification shortens the average transport paths between source and target data, leading to significantly faster training convergence and improved sample quality at lower numbers of function evaluations (NFEs) compared to baselines on ImageNet-64 and MS-COCO.
Key insights
- Initializing samples from a prior distribution that is, on average, closer to the target conditional mode significantly reduces the "transport cost" and straightens flow trajectories, leading to lower global truncation errors in numerical integration.
- Explicitly leveraging conditioning information to shape the prior distribution allows for more efficient and direct mappings between the source and target distributions.
- An informative conditional prior distribution can enable faster training convergence and generate high-quality samples with substantially fewer numerical integration steps (NFEs) during inference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.