Year: 2025
Type: conference-paper
Abstract: In recent years, with the high speed development of generative models, the quality of the image generated by text-to-image model has been improved greatly. However, in actual production environment, the human's body in the image generated by current model still has big flaws. Distorted limbs and deformed facial features in the generated human images are one of the main problems faced by the text-to-image model. Despite the importance of this problem, there is still no dedicated benchmark for evaluating the quality of generated human limbs. To solve this problem, we first constructed a dataset related to the body mass of the people in the image using generative models with different performances. Then, based on this dataset, we trained a scoring model called BodyMetric based on ImageReward. After training the scoring model, we can theoretically perform quality assessment on larger data. We use the scores provided by the scoring model and the DPO reinforcement learning strategy on the sdxl model to improve sdxl's ability to generate human images. A large number of experiments have shown that our scoring model can give scores that are consistent with human preferences for the body quality of the people in the generated images. At the same time, other experiments have shown that the generative model trained based on our scoring model has a good performance in character generation.
Author: Yuehao Zhao
Institution: Beijing Institute of Technology
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4411887362