SoPo: Text-to-Motion Generation Using Semi-Online Preference Optimization

Xiaofeng Tan, Hongsong Wang, Xin Geng, Pan Zhou · 2024

Read SoPo: Text-to-Motion Generation Using Semi-Online Preference Optimization by Xiaofeng Tan, Hongsong Wang, Xin Geng, Pan Zhou (2024) 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-2412-05095.txt.

Description

Xiaofeng Tan and three colleagues fine-tune text-to-motion generators so the movements they produce match what people actually prefer, in a 2024 paper. Turning words into believable motion matters for creative work, but models often drift into stiff or inconsistent moves.

The team first compares two training styles for preference learning and finds a flaw in each: one overfits its fixed data, the other samples with bias. Their method, SoPo, mixes the two, pairing fresh disliked samples with stored liked ones. On two established motion models it beat rival alignment methods on quality measures, and even passed a leading model on precision scores. Visual checks backed the gains.

Catalog record

CREATORXiaofeng Tan, Hongsong Wang, Xin Geng, Pan Zhou
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE117 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-05095

Browse the full library · About the Tower of Babel