codec integration

This work addresses a critical limitation in existing neural network compression methods: their neglect of cross-layer redundancy arising from functional symmetries, such as permutation invariance among hidden units and attention heads. To overcome this, the authors propose a novel framework that aligns symmetric blocks across layers via motion compensation, transforming weight sequences into predictable structures. The approach introduces a lightweight per-layer predictor, a rate–distortion-optimized entropy model, and a keyframe scheduling mechanism to efficiently encode quantized residuals. During decoding, inverse alignment enables rapid weight reconstruction. By explicitly modeling cross-layer alignment for the first time, the method achieves substantial improvements over state-of-the-art quantization and learned compression techniques on Transformer-based language modeling and vision classification tasks, significantly advancing the rate–accuracy Pareto frontier while preserving inference speed.

To address the bottlenecks of high bandwidth consumption, significant latency, and weak privacy protection in cloud-edge collaborative inference for edge-intelligent vision deployment, this paper proposes FCTM—the first standardized intermediate feature coding/decoding framework tailored for machine vision tasks, aligned with the MPEG Feature Coding Model (FCM) standard. Methodologically, FCTM integrates task-aware feature importance modeling, joint optimization of quantization and entropy coding, and lightweight inter-frame and inter-channel redundancy elimination, achieving a unified trade-off between semantic fidelity and high compression ratio. Evaluated across detection, segmentation, and recognition tasks, FCTM achieves an average bitrate reduction of 85.14% with negligible accuracy degradation (<0.3% mAP/mIoU), enabling real-time inference. End-to-end accuracy is fully preserved, and the framework has been officially validated by MPEG.

To address low data transmission efficiency and high terminal power consumption in edge-cloud collaborative AI inference, this paper proposes a novel edge-cloud co-inference framework based on intermediate neural feature encoding and compression. The method introduces intermediate-layer neural features as standardized coding units into the MPEG-AI international standard—Feature Coding for Machines (FCM)—enabling lossless, feature-level collaborative inference. It integrates lightweight feature extraction, adaptive quantization-based compression, and joint edge-cloud scheduling. Experimental results demonstrate that, compared to conventional remote inference, the framework reduces transmission bitrate by 75.90% while preserving model accuracy without degradation. It significantly improves inference latency and terminal energy efficiency, offering an efficient and practical solution for deploying large-scale models on resource-constrained, low-power devices.

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.

This work addresses the challenge of balancing compression efficiency and computational complexity in practical deployments of intelligent video coding. We propose an end-to-end low-complexity coding framework tailored to standardized common test conditions, integrated into the AVS-EEM platform. By leveraging a customized neural network architecture, efficient training strategies, and inference optimization techniques—all while strictly adhering to conventional coding common test conditions—the proposed approach substantially reduces computational overhead. After more than two years of iterative development, the latest model significantly outperforms the AVS3 reference software in compression performance under identical test conditions, marking a critical step toward the standardization and practical adoption of end-to-end intelligent video coding.

Existing end-to-end video compression methods lack content adaptivity, leading to inaccurate motion compensation and imprecise alignment. To address this, we propose a content-adaptive motion alignment framework with three key innovations: (1) an optical-flow-guided coarse-to-fine deformable warping module augmented with mask modulation to enhance deformation accuracy; (2) a reference-frame-quality-driven hierarchical distortion-weighting strategy to improve reconstruction fidelity; and (3) a motion-magnitude-aware, training-free dynamic frame downsampling scheme for resolution-adaptive optimization. All components are jointly optimized in an end-to-end manner. Evaluated on standard benchmarks, our method achieves a 24.95% BD-rate reduction (in PSNR) over the DCVC-TCM baseline, significantly outperforming both our reimplementation of DCVC-DC and the HEVC reference software HM-16.25.

This work addresses the storage, transmission, and deployment challenges posed by the massive parameter counts of large language models by introducing, for the first time, a systematic application of modern video compression techniques to model weight quantization. The proposed method integrates affine quantization with advanced video coding standards such as VVC/H.266, naturally aligning with the structural properties of weight matrices without requiring fine-tuning or calibration data. It demonstrates strong generalization across diverse tensor types. Experimental results on the LLaMA-3-8B model at 2-bit compression show a more than 1.5× reduction in perplexity and a 21% improvement in downstream task accuracy compared to existing approaches, substantiating the method’s efficiency, robustness, and broad applicability.

Existing image compression methods struggle to simultaneously satisfy the requirements of human vision and diverse machine vision tasks within a single model, and they often lack dynamic adaptation to the semantic importance and complexity of different image regions. To address this, this work proposes MoECodec—a token-aware Mixture-of-Experts image compression framework that replaces conventional feed-forward network (FFN) layers in a Transformer architecture with a dynamic, token-level expert mixture mechanism. The framework employs a content- and task-aware routing strategy, stabilizes expert assignment via spatial total variation regularization, and introduces a lightweight Group Shuffle MLP as the expert structure to enable efficient and coherent computational resource allocation. Experiments demonstrate that MoECodec significantly outperforms existing approaches in both image reconstruction quality and performance across multiple downstream machine vision tasks, confirming the effectiveness and generalization capability of a unified multi-task compression model.

General-purpose compression algorithms often struggle to simultaneously achieve high compression ratios and high throughput with low overhead, whereas specialized compressors, while offering superior performance, incur high development and maintenance costs and suffer from limited applicability. This work proposes a novel “graph-based” compression framework that, for the first time, models the compression process as a modular composition of encoders and decoders represented by a directed acyclic graph. By integrating a self-describing format with a universal decoder (OpenZL), the framework unifies the generality of generic methods with the performance of specialized ones. The approach substantially reduces the cost of developing and deploying domain-specific compressors, outperforming mainstream general-purpose compressors in both compression ratio and speed across multiple real-world datasets. It remains competitive with deep learning–based methods while operating orders of magnitude faster, and internal adoption at Meta has reduced development cycles from months to days.

This study presents the first empirical evaluation of entropy-conserving binarization (ECB) within a real-world CABAC framework, assessing both compression efficiency and computational overhead. Using a custom M-coder-based CABAC encoder, the authors integrated ECB alongside UEG, single-context Huffman, and HuffmanPos, conducting 2,480 bit-exact round-trip tests across synthetic data, procedurally generated images, and the Kodak dataset. Results demonstrate that ECB consistently outperforms single-context Huffman across all quantization parameters, achieving rate savings of 0.031–0.113 bits per symbol, while HuffmanPos surpasses other methods in 12 out of 15 source units. The primary driver of rate differences is attributed to context assignment rather than binarization length. To address ECB’s 7–10× higher decoding latency, the work proposes a single-pass interleaved decoding scheme to mitigate delay.