TP-ViT: truncated uniform-log2 quantizer and progressive bit-decline reconstruction for vision Transformer quantization

Abstract

Vision Transformers (ViTs) have achieved remarkable success across various artificial intelligence-based computer vision applications. However, their demanding computational and memory requirements pose significant challenges for deployment on resource-constrained edge devices. Although post-training quantization (PTQ) provides a promising solution by reducing model precision with minimal calibration data, aggressive low-bit quantization typically leads to substantial performance degradation. To address this challenge, we present the truncated uniform-log2 quantizer and progressive bit-decline reconstruction method for vision Transformer quantization (TP-ViT). It is an innovative PTQ framework specifically designed for ViTs, featuring two key technical contributions: (1) truncated uniform-log2 quantizer, a novel quantization approach which effectively handles outlier values in post-Softmax activations, significantly reducing quantization errors; (2) bit-decline optimization strategy, which employs transition weights to gradually reduce bit precision while maintaining model performance under extreme quantization conditions. Comprehensive experiments on image classification, object detection, and instance segmentation tasks demonstrate TP-ViT's superior performance compared to state-of-the-art PTQ methods, particularly in challenging 3-bit quantization scenarios. Our framework achieves a notable 6.18 percentage points improvement in top-1 accuracy for ViT-small under 3-bit quantization. These results validate TP-ViT's robustness and general applicability, paving the way for more efficient deployment of ViT models in computer vision applications on edge hardware.

1. Introduction

The vision Transformer (ViT) architecture (Dosovitskiy et al., 2021) has revolutionized computer vision by leveraging its innovative self-attention mechanism, establishing itself as a formidable alternative to traditional convolutional neural networks (CNNs). ViT-based models have achieved state-of-the-art (SOTA) performance across diverse artificial intelligence (AI)-driven engineering applications, including but not limited to image classification (Mahmood et al., 2024; Zhang ZC et al., 2024), object detection (Chen et al., 2024; Gao et al., 2024), semantic segmentation (Li MH et al., 2024; Zheng et al., 2025), and image processing (Tian et al., 2024; Xia et al., 2025). The key advantage of ViT lies in its ability to capture global contextual information, which has significantly advanced vision-based automation and the development of intelligent systems.

Despite their remarkable performance, Transformer-based models like ViT face significant deployment challenges in real-world engineering applications due to their large model size and high computational demands. The self-attention mechanism, while highly effective in capturing global dependencies, introduces substantial memory and processing overhead, particularly when processing high-resolution images, where computational complexity grows quadratically with input size (Zamir et al., 2022). These limitations pose critical obstacles for real-time deployment on resource-constrained platforms, which often operate under strict power and latency constraints.

To address these challenges, neural network quantization has emerged as a promising solution, effectively reducing model size and computational requirements while maintaining acceptable accuracy. By compressing model parameters into lower-bit representations, quantization significantly enhances the feasibility of deploying ViT models in practical and large-scale applications. There are two primary types of quantization in neural networks: quantization-aware training (QAT) (Kim et al., 2022; Nagel et al., 2022; Zhong et al., 2022; Liu SY et al., 2023) and post-training quantization (PTQ) (Li YH et al., 2021; Jiang YF et al., 2025; Zhong et al., 2026). QAT incorporates quantization constraints during the training phase, requiring retraining on labeled datasets to mitigate accuracy loss. Although effective, this approach imposes significant computational burdens, particularly for large Transformer models, making it resource-intensive and time-consuming. In contrast, PTQ operates on pre-trained models, requiring only a small amount of unlabeled calibration data to perform quantization. This method.