Laplacian Multi-scale Flow Matching for Generative Modeling

Zelin Zhao & Petr Molodyk & Haotian Xue & Yongxin Chen · 2026

Read Laplacian Multi-scale Flow Matching for Generative Modeling by Zelin Zhao & Petr Molodyk & Haotian Xue & Yongxin Chen (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-19461.txt.

Description

Flow matching builds images by learning flows between distributions, but handling multiple scales usually means cascaded steps with renoising between them. Zelin Zhao and three co-authors present LapFlow, Laplacian multiscale flow matching, which splits images into Laplacian pyramid residuals and handles scales in parallel. A mixture-of-transformers architecture with causal attention handles the scales together.

Generating multi-scale representations in parallel removes the need for bridging processes between scales. The design improves generation quality, speeds up sampling, and helps flow matching methods scale. On CelebA-HQ and ImageNet, LapFlow reaches superior sample quality with fewer computational operations and faster inference than single-scale and multi-scale baselines, a win on both quality and cost.

Catalog record

CREATORZelin Zhao & Petr Molodyk & Haotian Xue & Yongxin Chen
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE63 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-19461

Browse the full library · About the Tower of Babel