residual vector quantization

Existing vector quantization methods are constrained by static codebooks, limiting their ability to adapt to the heterogeneous geometric structures of data, while dynamic quantizers often suffer from inefficient serial decoding. This work proposes RQ-MoE, a novel framework that introduces the mixture-of-experts (MoE) mechanism into residual quantization for the first time. By employing a two-layer MoE module within a dual-stream architecture, RQ-MoE enables input-adaptive dynamic codebook construction and decouples token generation from quantization to facilitate parallel decoding. The framework unifies standard residual quantization and QINCo as special cases and provides design guidelines for expert dimensionality. Experiments demonstrate that RQ-MoE achieves state-of-the-art or comparable performance in reconstruction and retrieval tasks while delivering 6–14× faster decoding speeds than existing approaches.

To address the bandwidth bottleneck in AI model split inference caused by intermediate feature transmission, this paper proposes a lightweight, lossless feature compression method based on Z-score normalization. The core innovation lies in the first integration of Z-score standardization into a feature compression framework to explicitly preserve global statistical properties—namely, mean and variance—enabling an end-to-end differentiable encoding architecture. Unlike the conventional scaling scheme in the MPEG FCM standard draft, our approach yields a more compact and hardware-efficient implementation. Extensive evaluation across multiple vision tasks demonstrates an average bitrate reduction of 17.09%, with up to 65.69% savings in object tracking, while maintaining zero accuracy degradation in downstream tasks.

Current neural audio codecs employ quantization schemes—such as residual vector quantization (RVQ), vector quantization (VQ), and finite scalar quantization (FSQ)—that suffer from limited geometric modeling capacity in latent space, resulting in weak correlation capture among features, low codebook utilization, and high token rates. To address this, we propose Q2D2, a geometry-aware audio compression framework that, for the first time, jointly quantizes feature pairs onto structured 2D grids (hexagonal, rhombic, or rectangular), implicitly constructing efficient codebooks. This approach abandons the oversimplified manifold assumptions of scalar or vector quantization, enhancing geometric consistency and collaborative feature representation while preserving reconstruction fidelity. Experiments on speech reconstruction demonstrate that Q2D2 matches or surpasses state-of-the-art models in both objective and subjective quality metrics, achieves significantly higher codebook utilization, and validates—through ablation—the critical role of 2D grid-based quantization.

本文提出了一种几何感知的双曲残差量化方法，解决了双曲空间中残差聚合和梯度估计的几何不一致性问题，适用于层次化数据表示。

Existing video compression methods are primarily optimized for human visual perception and thus often fail to preserve semantic information critical for machine vision tasks. To address this, this paper proposes a machine-vision-oriented neural preprocessing framework. It introduces a learnable preprocessor prior to standard video encoding and pioneers a differentiable virtual codec, enabling end-to-end joint optimization of preprocessing and conventional encoders (e.g., H.264/AVC) without modifying codec standards. A rate–distortion–task loss jointly optimizes bit rate, reconstruction fidelity, and downstream task performance—including object detection and action recognition. Experiments demonstrate that the framework reduces average bit rate by over 15% while maintaining or even improving task accuracy, significantly enhancing semantic fidelity and utility of compressed video for machine vision applications.

Discrete audio representations in speech language models often degrade downstream task performance due to information loss. To address this, this work proposes a hybrid discrete-continuous modeling approach that jointly represents speech using temporally compressed discrete tokens and dimensionality-reduced continuous residuals. The method introduces a novel encoder-decoder architecture incorporating fusion-focused modulation and a hybrid Transformer design, enabling autoregressive inference in the discrete domain while simultaneously leveraging non-autoregressive prediction and continuous residual upsampling. This approach achieves the first effective integration of discrete and continuous representations, substantially reducing the number of autoregressive steps while preserving speaker characteristics and fine-grained acoustic details. Experimental results demonstrate clear performance gains over purely discrete baselines.

This study addresses the limitation that quantization and generative residuals impose on reconstruction quality in autoregressive image coding by proposing the ResARC framework. This method introduces a novel explicit dual residual compensation mechanism, leveraging diffusion models to compensate for quantization residuals while compressing and transmitting generative residuals. By integrating diffusion transformers with learned codecs, ResARC achieves high-fidelity context-based reconstruction without requiring additional side information. Experimental results demonstrate that ResARC significantly enhances distributional fidelity at ultra-low bitrates while maintaining highly competitive perceptual similarity.

This study addresses the optimization of high-precision coordinate retention strategies in rotation-based quantization. We propose a method that jointly optimizes the number and positions of retained coordinates before and after rotation to minimize quantization error under a fixed bit budget. Theoretically, we prove that retaining the top-k coordinates prior to rotation minimizes the upper bound of the quantization error, thereby reducing a complex combinatorial search to the optimization of a scalar k, for which a fast parallel selection algorithm is designed. Combined with random rotation preprocessing and offline codebook optimization, our approach achieves efficient compression. Experiments demonstrate that the proposed method significantly improves the trade-off between reconstruction accuracy and storage efficiency across Gaussian modeling, nearest neighbor retrieval, KV cache compression, and activation quantization tasks.

该研究通过结合RVQ、U-Net像素空间残差校正和GAE误差界限保证的方法，解决了科学数据压缩中重建误差控制的问题。

This study addresses the suboptimality arising from the decoupling of pruning and quantization in SVD-based compression by proposing a unified co-optimization framework. Methodologically, it introduces a differentiable component-wise bit-width learning mechanism that automatically assigns zero bits to low-importance components, thereby achieving precise pruning. By integrating SVD decomposition with a joint optimization algorithm, the framework enables end-to-end synergy between pruning and quantization. Experimental results demonstrate that under extreme 1.61-bit compression, the proposed approach significantly outperforms two-stage baseline methods. This work establishes a new paradigm for highly efficient compression of large models.