OneVision: An End-to-End Generative Framework for Multi-view E-commerce Vision Search

TL;DR

OneVision presents a unified end-to-end generative framework for e-commerce vision search that explicitly handles multi-view object discrepancies and aligns retrieval with user preferences. It introduces Vision-aligned Residual Quantization (VRQ) to produce consistent yet distinctive semantic IDs, and a multi-stage generative training pipeline (Pretraining, SFT, DPO) to fuse visual semantics with personalized signals. Dynamic token pruning further enhances online efficiency, enabling real-time inference with minimal accuracy loss. Across offline evaluations and large-scale online A/B tests, OneVision matches or surpasses traditional multi-stage architectures while delivering significant gains in CTR, CVR, and orders, demonstrating a practical, scalable shift toward semantic ID–driven generative retrieval in industrial settings.

Abstract

Traditional vision search, similar to search and recommendation systems, follows the multi-stage cascading architecture (MCA) paradigm to balance efficiency and conversion. Specifically, the query image undergoes feature extraction, recall, pre-ranking, and ranking stages, ultimately presenting the user with semantically similar products that meet their preferences. This multi-view representation discrepancy of the same object in the query and the optimization objective collide across these stages, making it difficult to achieve Pareto optimality in both user experience and conversion. In this paper, an end-to-end generative framework, OneVision, is proposed to address these problems. OneVision builds on VRQ, a vision-aligned residual quantization encoding, which can align the vastly different representations of an object across multiple viewpoints while preserving the distinctive features of each product as much as possible. Then a multi-stage semantic alignment scheme is adopted to maintain strong visual similarity priors while effectively incorporating user-specific information for personalized preference generation. In offline evaluations, OneVision performs on par with online MCA, while improving inference efficiency by 21% through dynamic pruning. In A/B tests, it achieves significant online improvements: +2.15% item CTR, +2.27% CVR, and +3.12% order volume. These results demonstrate that a semantic ID centric, generative architecture can unify retrieval and personalization while simplifying the serving pathway.

Figures (7)

• Figure 1: Comparison between the proposed end-to-end generative framework and the traditional multi-stage cascading architecture in e-commerce vision search.
• Figure 2: The major vision search scenarios in e-commerce on the Kuaishou Platform.
• Figure 3: Overall Framework of OneVision. The first row illustrates our VRQ codebook training and encoding process. The second row shows our multi-stage generative training pipeline, including Pretraining, SFT, and DPO. The third row presents our dynamic pruning and end-to-end inference process.
• Figure 4: Impact of Different Codebook Depth $L$ and Size $K$.
• Figure 5: Impact of Visual Token Number on Retrieval Accuracy (HR@4, MRR@4) and System Efficiency.