面向文本引导图像生成任务的生成对抗网络架构搜索方法
By constructing a text-conditional generative adversarial network supernet and using a multi-objective evolutionary algorithm to search for the generator architecture, combined with Logistic adversarial loss and instance noise, the problem of insufficient generation quality and semantic consistency of generative adversarial networks in text-guided image generation tasks is solved, and more stable generation results are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF INFORMATION SCI & TECH
- Filing Date
- 2026-05-08
- Publication Date
- 2026-07-17
AI Technical Summary
Existing generative adversarial networks lack systematic structural optimization in text-guided image generation tasks, which leads to the generator sacrificing low-level visual quality when the image-text alignment loss is too strong in the early stage of training. Furthermore, the training environment is inconsistent between the search and retraining stages, making it difficult to guarantee generation quality and semantic consistency.
A text conditional generative adversarial network supernet is constructed, employing weight-sharing training and a multi-objective evolutionary algorithm to search for the generator architecture. It combines Logistic adversarial loss, Lazy R1 gradient regularization, and instance noise, and uses a pre-trained image-text alignment model for semantic constraints. A consistent training strategy is maintained during the retraining phase to optimize the generator and discriminator parameters.
It improves the visual quality and semantic consistency of generated images, reduces the risk of performance degradation during the search and retraining phases, and significantly enhances the stability and reproducibility of generative adversarial networks in text-guided image generation tasks.
Smart Images
Figure CN122133725B_ABST