Imbalance in Balance: Online Concept Balancing in Generation Models
Yukai Shi, Jiarong Ou, Rui Chen, Haotian Yang, Jiahao Wang, Xin Tao, Pengfei Wan, Di Zhang, Kun Gai · 2025
Read Imbalance in Balance: Online Concept Balancing in Generation Models by Yukai Shi, Jiarong Ou, Rui Chen, Haotian Yang, Jiahao Wang, Xin Tao, Pengfei Wan, Di Zhang, Kun Gai (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-2507-13345.txt.
Description
Yukai Shi and co-authors tackle, in a 2025 paper, a weak spot of image generation models: complex concepts in one prompt often come out wrong or unstable. A scene that asks for several things at once may drop one or swap traits between them. Why this fails was not well mapped.
The team ran designed tests to trace the causal factors behind poor concept responses. They then added a loss function, called IMBA loss, that equalizes the pull of each concept during training. The method works online, with no need to process datasets offline, and it takes only minimal code changes. On a new benchmark for hard concept mixes, plus two public test sets, the loss significantly improved the concept responses of baseline models. The scores stood up well next to rival fixes. Code is shared online.
Catalog record
| CREATOR | Yukai Shi, Jiarong Ou, Rui Chen, Haotian Yang, Jiahao Wang, Xin Tao, Pengfei Wan, Di Zhang, Kun Gai |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 42 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-13345