JPEG is a lossy compression technique which is the default format for all major digital cameras and estimated to be used for over 80% of all digital images on the web [1], [2]. During JPEG encoding, quantization is used to approximate frequency coefficients by a set of integers called the quantization table. We introduce OptQT as a new framework for the optimal design of JPEG quantization tables.
Block prediction and reference frame generation play a central role in contemporary video codecs and hold significant potential for improving compression efficiency. For motion vectors of varying reliability, directly averaging the bi-directional reference blocks introduces noise in block prediction. In advanced video codecs such as AV2, the generation of co-located reference frames (CLRFs) within the temporal interpolated prediction (TIP) module is also performed at the block level. However, the imprecise vectors in the motion field cause the associated block pairs to exhibit weak correlation in their high-frequency components, which is not addressed by existing methods. This creates an opportunity for the transformdomain temporal prediction (TDTP) approach. This work proposes an adaptive TDTP framework for block prediction. Independent linear predictors (LPs) are trained for each AC coefficient to estimate the corresponding temporal correlation. A novel adaptive backward updating scheme for linear predictors is employed, in which reconstructed blocks are used to update the statistics and refine LP parameters without introducing additional bitrate overhead. Simulation results on the TIP module in AV2 demonstrate that the proposed approach effectively adapts to frame statistics and enhances the efficiency of compound inter-prediction.
Despite strong generalization ability, large language models (LLMs) are resource-intensive to deploy. Post-training quantization (PTQ) effectively compresses pretrained LLMs without retraining. Although recent studies target extremely low-bit quantization, most methods rely on extra parameters and the effective bit-width usually exceeds 2 bits. We introduce a saliency-guided mixed-precision PTQ framework that achieves an average bit-width of 1.70 bits while maintaining competitive performance. The proposed method allocates bit-widths across weight groups under a strict total budget, guided by a cross-precision saliency metric. The key innovation is a learnable affine transformation consisting of rotation, scaling, and shifting to align quantization grids with local weight distributions to minimize error. Experiments show that the proposed method matches state-of-the-art low-bit PTQ performance, while reducing the average bit-width by approximately 0.55 bits in Table 1.
Data compression techniques significantly improve storage efficiency, bandwidth utilization, and energy efficiency, yet they introduce challenges for the rapid browsing and retrieval of valuable information within compressed data. Existing approaches achieve high-speed, lossless matching by exploiting the context-free property of automata. However, they are constrained by the recursive reference structures in compressed data, which necessitate state copying to ensure matching safety.
The explosive growth of global data has created an increasing demand for archival storage, where efficient compression is crucial to reduce capacity cost. However, existing archival compressors face a fundamental tradeoff: mainstream methods (e.g., ZSTD with level $21 / 22$) offer limited compression ratios, while context-mixing compressors (e.g., LPAQ) achieve higher ratios but are often too slow for practical use. Therefore, we present CuCM, a GPU-powered context-mixing compressor to overcome this tradeoff. By introducing pre-learning and batch update mechanisms, CuCM resolves the data dependencies inherent in the autoregressive modeling of contextmixing compressors. During compression, CuCM processes each predefined vector as a single unit. It utilizes the current model to predict the probability distribution for the entire vector, deferring model updates until the vector is fully processed. During decompression, CuCM employs an aggressive look-ahead strategy, preassuming bit values for context construction. It then retains only the outcomes of hypotheses that remain consistent with the actual decoded data. Experiments like figure 1 show that CuCM achieves up to $12.6 \times$ higher throughput than LPAQ while maintaining comparable compression ratios across both general-purpose and archival datasets.
This paper proposes HRAC, a novel compressor targeting high-entropy and highresolution astronomical data of both integer and floating-point types. HRAC leverages the distinct characteristics of high-frequency astronomical data across different dimensions: it reads the data along the dimension with the lowest variability and partitions the stream into blocks. For each block, HRAC calculates the mean value of the data after excluding the maximum and minimum, then applies differential coding using this mean against all data within the block to generate a residual sequence. This residual sequence is encoded using a prefix code that combines ideas from ExpGolomb and Elias gamma coding. When compressing floating-point data, the local smoothness assumption crucial for differential prediction is violated if two close values straddle zero. HRAC addresses this by selectively moving the sign bit after the exponent bits. For decompression, several parameters are required per block. Unlike the conventional approach of storing parameters in each block header, HRAC computes the optimal parameters from the previous block and reuses them for the next, eliminating their storage overhead. Experiments on multiple datasets demonstrate that HRAC delivers superior overall performance compared with other compressors, as shown in Fig. 1.
3D Gaussian splats (3DGS) have rapidly gained traction as a 3D scene representation technique that enables efficient real-time rendering and highfidelity novel view synthesis. This paper reports on ongoing MPEG GSC efforts conducted jointly by the MPEG Video Coding group (WG 4) and the Coding of 3D Graphics and Haptics group (WG 7) to define a practical and interoperable compression framework for 3DGS. MPEG GSC is planning GSC standardization with short-term and long-term timelines to address market requirements. The short-term objective is to standardize coding tools that build on proven MPEG ecosystems while introducing only the minimal set of extensions, syntax, and processing required for INRIA-3DGS format (referred to in MPEG as I-3DGS). In the long-term, MPEG is also investigating broader alternatives for 3DGS representation and compression, including approaches that integrate training during compression, collectively referred to as Alternative-3DGS (A-3DGS). More specifically, this paper focuses on the I-3DGS and explores both geometrybased and video-based coding frameworks within MPEG GSC.
With the rapid advancement of Augmented Reality (AR) and autonomous driving, the storage and transmission of high-density point clouds have become critical challenges. To address this, we propose HMD-PCGC, a novel end-to-end Point Cloud Geometry Compression framework that synergizes a Mamba-Diffusion architecture with an adaptive hybrid representation learning mechanism. Unlike traditional approaches that rely on single-modality inputs, our encoder integrates three complementary feature types—point-based, voxel-based, and implicit features—through a Cross-Representation Attention module. This design allows for the comprehensive capture of local geometric details and complex topologies. To effectively model long-range dependencies within the latent space without the computational bottleneck of Transformers, we employ a Mamba block (State Space Model) for sequential processing, coupled with a scorer-guided adaptive quantization module that dynamically allocates bit-rate based on feature significance. For the decoding stage, we utilize a conditional Diffusion Model that generates high-fidelity point clouds by reversing a noising process, conditioned on the compressed latent representation. Extensive experiments on the MPEG 8iVFB and MVUB datasets demonstrate that HMD-PCGC achieves significant improvements in compression efficiency. Specifically, our method yields average BD-Rate gains exceeding 90% against G-PCC (octree), 80% against G-PCC (trisoup), and 40% against V-PCC, while also outperforming several recent learning-based frameworks in both rate-distortion efficiency and visual reconstruction quality.
In this paper, we propose a lightweight multiscale reference frame generation network for VVC inter coding, named LMRFG. Unlike the previous work [1], LMRFG does not employ the high-performance operation point (HOP) network as preprocessing for frame enhancement. LMRFG replaces the bidirectional motion estimation network with a dualbranch coordinated attention motion estimator (DBCA-ME), which integrates X-axis and Y-axis optical flows for accurate optical flow estimation. Moreover, LMRFG uses depthwise over-parameterized convolutional layer (DO-Conv) to reduce model complexity and minimize bitstreams while maintaining video quality. LMRFG adopts a quantization parameter (QP) distance-based training strategy that takes compressed data at higher QP as input and compressed data at lower QP as label for training, thus addressing the imbalanced QP gap between the compressed input and its uncompressed label. LMRFG is embedded between DPB and RPL to generate a new reference frame and replace the original reference frame in RPL. As shown in Table 1, LMRFG achieves average BD-rate gains of {RA: 4.31% (Y), 6.54% (U), 7.05% (V)} and $\{\text{LDB}: 3.81 \%(\mathrm{Y}), 8.90 \%(\mathrm{U}), 8.77 \%(\mathrm{V})\}$ over the VTM_11.0-NNVC_10.0 (NN-tools ON) anchor, achieving state-of-the-art performance.
Current machine learning research on image compression mainly focuses on developing neural network architectures and preparing training datasets with compressed input and raw uncompressed labels. This study originates from a fundamental question of whether using raw uncompressed data as labels is the best approach to compression models. To address this question, we employ less compressed data as labels, rather than raw uncompressed data, in training the compression model. Moreover, we investigate how learning incrementally between compressed datasets can produce higher quality images than using raw uncompressed images (ground truth) as labels. Based on the results, we propose Compression Artifact-Aware Incremental Machine Learning for JPEG image compression, called CAIML. For JPEG compression artifact removal, we adopt ARCNN [1], which has the simplest network architecture and allows for a clear assessment of performance changes caused by different label settings. In CAIML, the label dataset is constructed with JPEG quality settings in consistent incremental steps, enabling ARCNN to learn compression artifact reduction incrementally. Table 1 shows the results obtained using JPEG, ARCNN, CAIML with increments of 10 (CAIML-INC10), and CAIML with increments of 20 (CAIML-INC20) starting at a quality factor of $q=10$. Both CAIML variants outperform the original ARCNN. Among them, CAIML-INC10 achieves the highest PSNR, indicating the lowest distortion. The results indicate that when neural networks incrementally learn from less compressed data, they achieve higher quality than learning from the raw uncompressed data (ground truth). CAIML is expected to be applicable to various neural network-based compression models for both images and videos.11This work was conducted during Hanna Jung's internship at Xidian University.
Existing image compression models often lack personalization capabilities, treating all image regions equally and failing to meet the compression needs of different users for specific Regions of Interest (ROI). To address this challenge, we propose an innovative variable rate image compression framework that achieves user-centric dynamic compression by introducing visual in-context learning. Our method extracts cross-image semantics from user-provided visual examples to understand their intent. This semantic information is then converted into visual semantic query tokens and spatial masks to effectively guide the bit allocation of the compression model. Furthermore, we design a novel Semantic Spatial Control Block (SSCB) to fully leverage these semantic and spatial cues, thereby achieving a balance between preserving user-specified details and overall image quality. Experimental results demonstrate that our method significantly improves performance on the ROI, achieving a 31.54 % BD-Rate reduction and a 2.7479 dB BD-PSNR gain over the baseline model.
In the AVS3 reference software HPM, intra prediction filtering (IPF) is employed in inter prediction filtering (Interpf). However, since IPF was originally designed for spatial boundary pixels, its direct application to Interpf is suboptimal. To address this issue, we propose a Temporal IPF technique (Temp-IPF). Temp-IPF compensates for the lack of temporal correlation when IPF is used in inter prediction, thereby improving the performance of inter prediction filtering in AVS3. Concurrently, we have redesigned the filtering parameters of IPF. Experimental results on the AVS3 720p test sequences show that Temp-IPF and the adjusted-IPF (with the redesigned parameters) achieve average BD-rate reductions of 0.025% and 0.13%, respectively.
In the latest international video coding standard, Versatile Video Coding, merge mode utilizes spatial or temporal adjacent motion information as motion vector predictors to reduce the signaling overhead. However, only one merge candidate's motion information is selected from previously coded blocks for motion compensation and prediction. To improve the prediction efficiency, a dual merge mode which inherits two candidates' motion information from previously coded blocks is proposed. In the proposed method, motion compensation is executed utilizing the motion information from both merge candidates to generate two predictions, and the two predictions are ultimately combined with equal weight to derive the final predicted results. Both of the merge candidates can be uni-prediction or bi-prediction, requiring up to four motion compensations for the proposed dual merge mode. Experimental results show that the proposed dual merge mode achieves an average coding gain of 0.03 % over ECM-14.0 with negligible complexity increase, and the extended ablation studies verify the effectiveness of the proposed method.
3D Gaussian splatting enables real-time, photo-realistic scene rendering but is challenged by redundant, high-dimensional primitive data. To make immersive AR/VR and volumetric communication practical, compression must address both temporal and attribute redundancy. We introduce a streamlined inter-prediction coding pipeline that predicts each P-frame's color and geometry attributes from a nearby reference using a lightweight hybrid predictor, combining bilateral filtering and K-nearest-neighbor feature transfer, and encodes only residuals. Our approach targets the most rate-critical channels by applying codebook-based vector quantization to spherical harmonic components, while retaining efficient scalar quantization for other attributes. All quantized streams and codebooks are entropy coded. The proposed method achieves an average 89.1 % reduction in BD-rate and a 12.39 dB increase in BD-PSNR over GPCCv1, and outperforms previous SOTA interGS by 18.6 % in BD-rate and 1.81 dB in BD-PSNR. With practical decoding speeds (0.605 seconds per frame), interGS-Lite preserves real-time rendering and has potential for streaming and AR/VR applications, delivering consistent bitrate reductions with high visual quality.
We show that the enumeration of unbordered words can be done with amortized $O(1)$ time per reported word via a compressed representation of the reported words. More precisely, we represent unbordered words as cyclic shifts of Lyndon words, and enumerate all words corresponding to the same Lyndon word together. While storing explicitly a Lyndon word of length $n$ with $k$ unbordered cyclic shifts requires $k n \lg \sigma$ bits, the space can be reduced to $n \lg \sigma+n$ bits by storing the Lyndon word explicitly and the cyclic shifts implicitly, where $\sigma$ denotes the size of an ordered finite alphabet. The time complexity stems from a linear-time algorithm to compute the border correlation function that specifies which cyclic shifts of a Lyndon word correspond to unbordered words. The experimental results show that the compression ratio is approximately $1 / n$ for large values of $n$ and $\sigma$ (when $n \geq 8$ and $\sigma \geq 10$). In addition, the algorithm can also be used to randomly generate at least one unbordered word in expected $O(n)$ time, which is faster than the $O\left(n^{2} / \lg n\right)$-time method in literature.
Film grain synthesis is a powerful tool that can significantly reduce the bitrate of a grainy video. It is typically used with noise removal before the compression, which can make banding more pronounced in the compressed video. When the synthesized grain is added, the banding can still be visible, even at mid QPs. This article describes three algorithms that can be used with the AV1/AV2 film grain synthesis to reduce visibility of underlying bands in the re-noised video. These changes to the film grain synthesis algorithm are computationally inexpensive and improve the perceptual video quality when banding is present.
We introduce Attractor Matching, a new framework for structural string comparison built upon the theory of string attractors. Given a pattern $x$ of length $m$ and one of its attractors $\Gamma_{x}$, the problem asks for all substrings $y[i. . i+m-1]$ of a text $y$ such that $\Gamma_{x}$ is also an attractor of $y[i. . i+m-1]$. Unlike classical notions of string matching, which rely on character equality or distance measures, attractor matching focuses on the structural properties that govern repetitiveness and compressibility. Our contribution is fourfold. First, we adapt the IsAttractor algorithm of Béal et al. by combining the DAWG with the slidingwindow technique of Blumer, enabling online attractor verification as the window advances over the text. Second, we reformulate the verification procedure on the Compressed DAWG (CDAWG), obtaining a more compact representation that preserves correctness. Third, we employ the sliding-window CDAWG method of Inenaga et al., which allows efficient attractor matching on sliding-window maintained CDAWGs with incremental updates. Finally, we introduce a relaxed variant, Attractor Matching with Mismatches, where the pattern attractor may be extended by at most $\rho$ additional positions, enabling structurally tolerant matching. This paradigm bridges compression and similarity, opening new directions for structure-aware pattern matching.
Coding of Gaussian splats has drawn the attention of academia and standardization bodies lately. A commonly used approach involves projecting explicit 3D attributes onto 2D planes of a video and compressing it with existing video coding standards. However, such an approach produces excessively high sample rates, resulting in a significant bottleneck that might make it unsuitable for existing hardware decoders if the number of splats is high. This paper presents a framework that utilizes an implicit representation, where Gaussian splat attributes are represented by compact feature planes. By reducing the dependency of video resolution on the number of splats, our approach significantly reduces the required sample rate, making it suitable for deployed devices. In addition, the proposed framework achieves a superior rate-distortion trade-off, providing high-fidelity reconstruction at low bitrates without the excessive sample rates associated with the conventional video-based anchor.
This paper proposes a novel neural B-frame video compression framework with hybrid implicit motion modeling. In our approach, implicit motion modeling replaces the rate-consuming yet less effective flow-based explicit motion modeling to improve overall RD performance. Specifically, an interpolated frame is first generated from the forward and backward reference frames to enrich the temporal priors. A Hybrid Temporal Prior Extractor (HTPE) is then introduced to exploit these priors, where a hybrid feature extractor combining Content-Aware Depthwise Separable Convolution (CADSC) and Linear Attention Duality (LAD) adaptively captures local and global temporal features, respectively. Finally, the enriched temporal prior features are leveraged in the main encoder/decoder to enable implicit motion modeling, and are further integrated into the entropy model to improve the accuracy of entropy estimation for the discrete latent representation.
Prefix-Free Parsing (PFP) enables efficient indexing of repetitive text collections through dictionary-based storage of distinct factors and the corresponding factor positions in the text. The parsing method of PFP differs from LZ78 compression as it separates factors at occurrences of predefined trigger strings, ensuring that the parse is prefix-free. This prefix-free property of PFP makes it suitable for compressed indexing applications such as the construction of the Burrows-Wheeler Transform. Recent work has also introduced a recursive PFP technique, in which the parse of the initial PFP is further compressed with an additional round of prefix-free parsing. In this paper, we provide a theoretical analysis of the minimal size of any prefix-free parse. We show that for a PFP of a string of length $n$, the sum of the lengths of the parse string and the number of nodes in the trie of dictionary strings must be at least $2 \sqrt{n}$. We generalize this result by demonstrating that if prefix-free parsing is applied recursively $k$ times, the combined size of the dictionary tries at each level and the final parse must be at least $(k+1) \cdot n^{1 /(k+1)}$. We also provide a family of strings and corresponding trigger strings that achieve this lower bound, demonstrating its tightness and a polynomially large separation between PFP and several other popular compressibility measures. Finally, we illustrate that optimal PFP - that is, with optimal trigger string selection - is robust against one-bit catastrophes for some strings where LZ78 is not. This demonstrates that the optimal PFP size cannot be guaranteed to be within a multiplicative ratio polynomially smaller than $n^{1 / 8}$ relative to the LZ78 size.
