Reproduce “Laplacian Multi-scale Flow Matching for Generative Modeling” 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
Zelin Zhao, Petr Molodyk, Haotian Xue, Yongxin Chen
Researchers at Georgia Institute of Technology developed Laplacian Multi-Scale Flow Matching (LapFlow), a generative model that employs a unified Mixture-of-Transformers architecture with Laplacian pyramid decomposition. LapFlow achieves improved image generation quality and computational efficiency, scaling effectively to high resolutions up to 1024x1024 while outperforming existing flow matching and diffusion models in FID scores and inference speed.
Key insights
- The unified MoT architecture with causal masking efficiently handles multi-scale processing within a single model, eliminating the need for separate networks or complex renoising steps.
- Laplacian pyramid decomposition effectively aids in modeling both global image structures and fine details, contributing to enhanced image quality and improved scalability.
- A linear noise schedule and specifically configured critical time points for progressive training empirically outperform generalized variance preserving (GVP) paths and other noise schedules.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.