Correlation-Weighted Multi-Reward Optimization for Compositional Generation

Jungmyung Wi, Hyunsoo Kim, Donghyun Kim † · 2026

Read Correlation-Weighted Multi-Reward Optimization for Compositional Generation by Jungmyung Wi, Hyunsoo Kim, Donghyun Kim † (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-2603-18528.txt.

Description

Wi, Kim, and Kim present correlation-weighted multi-reward optimization for compositional image generation. Text-to-image models often drop concepts from multi-concept prompts, giving only partial success. Competing concept rewards interfere when jointly optimized in a naive way. Their CMO framework splits prompts into concept groups and scores each with dedicated reward models. Correlation-based weights push hard, conflicting concepts most; SD3.5 and FLUX.1-dev improve on ConceptMix, GenEval 2, and T2I-CompBench.

Catalog record

CREATORJungmyung Wi, Hyunsoo Kim, Donghyun Kim †
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE61534
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.18528

Browse the full library · About the Tower of Babel