ARCHE: Autoregressive Residual Compression with Hyperprior and Excitation
Authors: Sofia Iliopoulou, Dimitris Ampeliotis, Athanassios Skodras
Organizations: Dept. of Electrical and Computer Engineering, University of Patras, Patras, Greece · Dept. of Digital Media and Communication, Ionian University, Argostoli, Greece
Recent progress in learning-based image compression has demonstrated that end-to-end optimization can substantially outperform traditional codecs by jointly learning compact latent representations and probabilistic entropy models. However, many existing approaches achieve high rate-distortion efficiency at the expense of increased computational cost and limited parallelism. This paper presents ARCHE - Autoregressive Residual Compression with Hyperprior and Excitation, an end-to-end learned image compression framework that balances modeling accuracy and computational efficiency. The proposed architecture unifies hierarchical, spatial, and channel-based priors within a single probabilistic framework, capturing both global and local dependencies in the latent representation of the image, while employing adaptive feature recalibration and residual refinement to enhance latent representation quality. Without relying on recurrent or transformer-based components, ARCHE attains state-of-the-art rate-distortion efficiency: it reduces the BD-Rate by approximately 48% relative to the commonly used benchmark model of Balle et al., 30% relative to the channel-wise autoregressive model of Minnen & Singh and 5% against the VVC Intra codec on the Kodak benchmark dataset. The framework maintains computational efficiency with 95M parameters and 222ms running time per image. Visual comparisons confirm sharper textures and improved color fidelity, particularly at lower bit rates, demonstrating that accurate entropy modeling can be achieved through efficient convolutional designs suitable for practical deployment.
Figures & tables
Method
Hyperprior
AR
CC
Context Type
Ballé et al. [ 22 ]
Yes
No
No
None
Minnen et al. [ 15 ]
Yes
Yes
No
Masked Convolution
Minnen & Singh [ 28 ]
Yes
No
Yes
None
Iliopoulou et al. [ 52 ]
Yes
Yes
Yes
LSTM
ARCHE (proposed)
Yes
Yes
Yes
Masked PixelCNN + SE
TABLE I: Comparison of Entropy Modelling Approaches
Fig. 1: ARCHE architecture with hyperprior, masked context model, latent residual prediction, and two example slices showing channel conditioning with SE blocks. Q: quantization; AE/AD: arithmetic encoder/decoder; μ , σ : distribution parameters.
Fig. 2: The architecture of the masked context model. Type A masked convolution zeroes out all pixels below and to the right of the center (including the center). Type B is the same but includes the center pixel itself.
Fig. 3: The architecture of the squeeze-and-excitation block. C denotes the number of channels and r is the reduction ratio.
Component
Encoder
Decoder
Main Transform
Conv 5 × 5 c192 s2, GDN
Deconv 5 × 5 c192 s2, IGDN
Conv 5 × 5 c192 s2, GDN
Deconv 5 × 5 c192 s2, IGDN
Conv 5 × 5 c192 s2, GDN
Deconv 5 × 5 c192 s2, IGDN
Conv 5 × 5 cl_d s2
Deconv 5 × 5 c3 s2
Hyper Transform
Conv 3 × 3 c320 s1, ReLU
Deconv 5 × 5 c192 s2, ReLU
Conv 5 × 5 c256 s2, ReLU
MaskConv 3 × 3 c192 s1
TABLE II: Network Architecture
Fig. 4: Rate-distortion curves on Kodak dataset. PSNR (left) and MS-SSIM (right) versus bit rate. ARCHE achieves superior performance across most operating points, surpassing VVC intra.
Fig. 5: Rate-distortion curves on Tecnick dataset. PSNR (left) and MS-SSIM (right) versus bit rate. ARCHE achieves superior performance across most operating points, surpassing VVC intra.
Method
BD-Rate vs Ballé
BD-Rate vs VVC
Minnen et al. [ 15 ]
-8.00
90.61
Minnen & Singh [ 28 ]
-16.28
63.55
WeConvene [ 35 ]
-6.92
92.47
Iliopoulou et al. [ 52 ]
-24.22
30.19
ARCHE (proposed)
-48.01
-5.61
TABLE III: BD-Rate Savings (%) vs Ballé [ 22 ] and VVC Intra on Kodak (PSNR)
Method
BD-Rate vs Ballé
BD-Rate vs VVC
Minnen et al. [ 15 ]
-8.81
79.04
Minnen & Singh [ 28 ]
-13.99
50.32
WeConvene [ 35 ]
-6.77
86.01
Iliopoulou et al. [ 52 ]
-22.71
33.59
ARCHE (proposed)
-44.89
-10.28
TABLE IV: BD-Rate Savings (%) vs Ballé [ 22 ] and VVC Intra on Tecnick (PSNR)
Fig. 6: BD-rate savings versus number of channel slices. Performance saturates around 10 slices.
Fig. 7: BD-rate savings of component ablations versus hyperprior baseline. Each module contributes complementary improvements.
Fig. 8: Visual comparison on Kodak at similar bit rates. ARCHE preserves sharper details and more natural colours.
Method
Parameters
Time (ms)
Ballé et al. [ 22 ]
11.7M
25
Minnen et al. [ 15 ]
95.8M
591
Minnen & Singh [ 28 ]
121.7M
249
Iliopoulou et al. [ 52 ]
124.3M
265
ARCHE (proposed)
95.4M
222
TABLE V: Model Complexity and Inference Time per image
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Fig. 9: The flow chart of the incorporation of the GMM into the ARCHE pipeline.
Fig. 10: Two-pass checkerboard decoding. The arrows show the four anchor neighbours giving spatial context to each non-anchor element.
Fig. 11: Rate-distortion curves on Kodak dataset. PSNR (left) and MS-SSIM (right) versus bit rate. ARCHE achieves superior performance across all operating points.
Fig. 12: Rate-distortion curves on Tecnick dataset. PSNR (left) and MS-SSIM (right) versus bit rate. ARCHE achieves superior performance across all operating points.
Entropy coding is widely used in typical learned image compression (LIC) that converts latents into a compact bitstream. However, entropy coding is typically sequential and becomes the coding latency bottleneck. To overcome it, we present Entropy-Coding Free Learned Image Compression (EF-LIC), a multi-rate framework that generates compact representation by removing statistical and correlation redundancy with low coding latency. First, we introduce unconstrained vector quantization and prove that its index distribution approaches the maximum-entropy bound, yielding minimal statistical redundancy. Second, we propose a context-conditioned autoregressive transform that directly reparameterizes the latents to reduce inter-dependency. Theoretical analysis shows that EF-LIC can remove correlation redundancy as effectively as typical LIC with entropy coding, leading to comparable compression performance. Experiments show EF-LIC achieves up to 67.86% bitrate reduction over MS-ILLM on Kodak with LPIPS. Ablation studies further show EF-LIC matches the compression performance of its entropy-coding based variant while achieving over 3× faster encoding and 5× faster decoding.
Hao Cao, Wenqi Guo, Zhijin Qin +1
Department of Electronic Engineering, Tsinghua University · Department of Automation, Tsinghua University · State Key Laboratory of Space Network and Communications +1
Progressive autoregressive image codecs provide an appealing paradigm for generative compression by quantizing continuous latents into discrete tokens, transmitting coarse-to-fine prefix tokens and generating the remaining suffix tokens at the decoder. However, their reconstruction quality is fundamentally limited by two residuals introduced along this pipeline: the quantization residual, arising from information loss during discrete tokenization, and the generation residual, resulting from imperfect autoregressive generation of the suffix tokens. To address these limitations, we introduce ResARC, a residual-aware autoregressive codec that explicitly compensates for both residuals at the decoder. Specifically, we generate the quantization residual with a diffusion transformer conditioned on the autoregressive decoding context, while requiring no additional side information. In parallel, we compute the generation residual at the encoder and employ a learned Generation Residual Codec to efficiently compress and transmit it for decoder-side correction. The recovered residuals are then integrated with the reconstructed latent representation and decoded through an adapted VAE decoder. Extensive experiments demonstrate that ResARC achieves competitive perceptual similarity while substantially improving distributional fidelity over leading generative codecs across the ultra-low bitrate regime. Code and models will be released soon.
Qin Yan, Ruixiao Dong, Yutao Xie +5
University of Science and Technology of China · Alibaba Group
Most learned image compression systems rely on a single latent representation combined with a hyperprior, which limits their ability to efficiently capture image structure across spatial scales. In this work, we propose a hierarchical latent representation to improve the efficiency of the entropy model. By using multiple latents at different scales, each with its own entropy model, we better capture the spatial structure of the latent representation. Our experiments show that this approach achieves a 17.9% BD-rate reduction over VVC on Kodak, demonstrating the effectiveness of multi-scale latent representations. Furthermore, the approach is orthogonal to other advances in learned image compression, making it a versatile addition to existing methods.
Jonas Brenig, Radu Timofte
Computer Vision Lab, CAIDAS & IFI, University of Würzburg, Germany