This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
Imbalance in Balance: Online Concept Balancing in Generation Models
Created by
Haebom
Author
Yukai Shi, Jiarong Ou, Rui Chen, Haotian Yang, Jiahao Wang, Xin Tao, Pengfei Wan, Di Zhang, Kun Gai
Outline
This paper addresses the problem that complex concept responses and combinations in visual generation tasks are unstable and error-prone. The authors explore the cause of this problem through elaborately designed experiments and propose a concept-wise equalization loss function (IMBA loss) to address it. The proposed method is an online approach that does not require offline dataset processing and minimizes code changes. On a newly proposed complex concept benchmark, Inert-CompBench, and two public test sets, our method significantly improves the concept response capability of the baseline model and achieves highly competitive results with minimal code changes.
Takeaways, Limitations
•
Takeaways:
◦
A novel approach to solving the problems of instability and errors that arise in the visual generation of complex concepts.
◦
Proposing an efficient concept-wise equalization loss function (IMBA loss) that can be applied online without offline processing.
◦
We experimentally demonstrate that the performance of existing models can be significantly improved with minimal code changes.
◦
Introducing a new complex concept benchmark, Inert-CompBench.
•
Limitations:
◦
Further verification of the generality and versatility of Inert-CompBench is needed.
◦
Further studies are needed to determine how well the proposed method generalizes to different visual generative models and tasks.
◦
Lack of detailed discussion on parameter optimization of IMBA loss.