LayerNorm-Aware Image Synthesis for Data-Free Quantization of Vision Transformers

Vision Transformers(ViTs) have recently achieved remarkable success across a wide range of computer vision tasks. However, their high model complexity poses significant challenges for deployment on resource-constrained devices. Quantization serves as an effective model compression technique, and data-free quantization-requiring no access to the original training datahas attracted increasing attention due to its ability to address data privacy and security concerns. Nevertheless, current data-free quantization methods face two major limitations: (1) the synthetic data often fails to align closely with the internal characteristics of the target model, and (2) the quantized models exhibit limited tolerance to distribution shifts. To overcome these challenges, we propose LNAQ-ViT(LayerNorm-Aware Image Synthesis for Data-Free Quantization of Vision Transformers), a novel framework designed for effective data-free quantization of ViTs. Our approach first employs an adaptive Nesterov optimizer, coupled with LayerNorm parameter importance guidance and Wasserstein distribution alignment, to generate synthetic images that can representative of the model’s internal features and is similar to the real image. We further introduce dynamic noise adaptation and adversarial boundary enhancement to produce critical samples that significantly improve the model’s robustness against distribution shifts. Extensive experiments demonstrate that LNAQ-ViT consistently achieves state-of-the-art accuracy across various Vision Transformer(ViT) architectures on image classification tasks, outperforming existing data-free quantization methods.