Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation

Dogyun Park, Taehoon Lee, Minseok Joo, Hyunwoo J. Kim · 2025

Read Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation by Dogyun Park, Taehoon Lee, Minseok Joo, Hyunwoo J. Kim (2025) 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-2510-21167.txt.

Description

Dogyun Park leads three co-authors on Blockwise Flow Matching, which splits the generative trajectory of flow matching models into time segments, each handled by a smaller specialized network. A single large network must capture every stage of the path from noise to data and is costly to evaluate repeatedly.

Distinct signal character across timesteps strains the one network design, and full model calls at every step inflate inference cost. Segment specialists can be lighter and better tuned to their slice of the trajectory. The framework aims at high quality image generation with materially lower compute per sample.

Catalog record

CREATORDogyun Park, Taehoon Lee, Minseok Joo, Hyunwoo J. Kim
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE60 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-21167

Browse the full library · About the Tower of Babel