VTONGuard: Automatic Detection and Authentication of AI-Generated Virtual Try-On Content

TL;DR

VTONGuard tackles the authenticity challenge of increasingly realistic AI-generated virtual try-on content by introducing a large-scale, real-vs-synthetic benchmark that spans diverse poses, backgrounds, and garment styles. The authors propose MiT-B2-MT, a multi-task detector that adds a segmentation head to learn boundary-aware features, achieving state-of-the-art performance on VTONGuard while keeping test-time cost unchanged thanks to training-only segmentation supervision. They systematically evaluate convolutional and transformer-based detectors across six VTON paradigms, revealing strong cross-subset generalization gaps and highlighting the ongoing need for paradigm-agnostic representations. The benchmark is further strengthened by IC-Light illumination harmonization and DIV2K-based background augmentation, enabling fair comparisons and facilitating robust, real-world deployment of VTON technologies, as evidenced by the improved performance of boundary-aware, high-resolution models like MiT-B2-MT. $L = L_{cls} + L_{seg}$ plays a central role in guiding the detector toward boundary artifacts during training.

Abstract

With the rapid advancement of generative AI, virtual try-on (VTON) systems are becoming increasingly common in e-commerce and digital entertainment. However, the growing realism of AI-generated try-on content raises pressing concerns about authenticity and responsible use. To address this, we present VTONGuard, a large-scale benchmark dataset containing over 775,000 real and synthetic try-on images. The dataset covers diverse real-world conditions, including variations in pose, background, and garment styles, and provides both authentic and manipulated examples. Based on this benchmark, we conduct a systematic evaluation of multiple detection paradigms under unified training and testing protocols. Our results reveal each method's strengths and weaknesses and highlight the persistent challenge of cross-paradigm generalization. To further advance detection, we design a multi-task framework that integrates auxiliary segmentation to enhance boundary-aware feature learning, achieving the best overall performance on VTONGuard. We expect this benchmark to enable fair comparisons, facilitate the development of more robust detection models, and promote the safe and responsible deployment of VTON technologies in practice.

Figures (2)

• Figure 1: Overall construction pipeline of the proposed VTONGuard benchmark, showing real and synthetic data collection, IC-Light harmonization, generative paradigm partitioning, and background augmentation with DIV2K for realistic VTON detection.
• Figure 2: Overall architecture of our MT design. The model adopts a transformer encoder for feature extraction and adds a segmentation branch alongside the classification head to jointly predict real/fake labels and human-clothing region masks, enabling boundary-aware feature learning.