QTok: Scalable 1D Visual Tokenization via Decoupled Querying
Abstract
One-dimensional (1D) image tokenizers have emerged as key enablers for efficient visual generation by compressing images into compact discrete sequences. However, existing approaches adopt a coupled architecture that concatenates learnable latent tokens with image patches at the input. This tight entanglement creates three critical bottlenecks: (1) structural incompatibility with powerful pretrained foundation models, (2) quadratic complexity that hinders scaling to larger token budgets, and (3) a representation bottleneck caused by single-scale feature extraction. To address these limitations, we propose QTok, a novel framework that introduces a paradigm shift to decoupled tokenization. We reformulate 1D tokenization as an ''Encode-then-Query'' process: a standard encoder first extracts a Hierarchical Multi-Level Memory to capture both global semantics and fine-grained details. Subsequently, a lightweight Query Attention Module distills this memory into compact latent tokens via cross-attention. This decoupling converts the scaling complexity from quadratic to linear and enables seamless integration with foundation models. Experiments on ImageNet-1K demonstrate that QTok establishes a new state-of-the-art for 1D tokenizers in generation fidelity, semantic consistency, and efficiency. Notably, our empirical results indicate that leveraging pretrained visual encoders achieves superior performance with only 50% of the standard training schedule. Code is available at https://github.com/Wenzhao-Xiang/QTok.