HP-GAN: Harnessing pretrained networks for GAN improvement with FakeTwins and discriminator consistency.

2026-01-01

Neural Networks

AI Summary

This work addresses the limitations of generative adversarial networks (GANs) in producing images with sufficient diversity and quality across varying data scales. To this end, the authors propose FakeTwins, a self-supervised loss mechanism, combined with a cross-architecture discriminator consistency strategy. By leveraging a pretrained network as a source of self-supervised signals, the method jointly trains multi-scale feature maps extracted from both CNN and Vision Transformer backbones, effectively integrating their complementary priors to enhance training stability and generalization. Evaluated on 17 diverse datasets spanning different image domains and data scales, the proposed approach consistently outperforms current state-of-the-art methods, achieving substantial improvements in Fréchet Inception Distance (FID) and generating images with markedly enhanced quality and diversity.

Technology Category

Computer Vision: Generative Adversarial Networks (GANs) for VisionMachine Learning: Deep Generative Models & AutoencodersNatural Language Processing: Generation

Application Category

Social Networks and Social Media: Generative AI / large language models and their impact on social systemsSearch and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAGWeb Mining and Content Analysis: Large pretrained models with web data

Problem

• Generative Adversarial Networks
• pretrained networks
• image synthesis
• self-supervised learning
• discriminator consistency

Innovation

• FakeTwins
• discriminator consistency
• pretrained networks
• self-supervised learning
• GAN improvement