ARCHE: Autoregressive Residual Compression with Hyperprior and Excitation¶

Conference: CVPR 2026 arXiv: 2603.10188 Code: GitHub Area: Learned Image Compression Keywords: Autoregressive entropy model, hyperprior, Squeeze-and-Excitation, residual prediction, rate-distortion optimization

TL;DR¶

A fully convolutional architecture that unifies hierarchical hyperprior, Masked PixelCNN spatial autoregression, channel-conditional modeling, and SE channel excitation — without relying on Transformers or recurrent components. With 95M parameters and a 222ms decoding time, it achieves a 48% BD-Rate reduction over the Ballé baseline and outperforms VVC Intra by 5.6%.

Background & Motivation¶

Background: End-to-end learned image compression has surpassed traditional codecs (JPEG/JPEG2000) by jointly optimizing analysis transforms, quantization, and entropy models for superior rate-distortion trade-offs. Transformer/attention architectures and hybrid entropy models have continually pushed the performance frontier.

Limitations of Prior Work: (1) Transformer/attention-based compression models incur heavy computation and slow inference, making deployment difficult; (2) ConvLSTM context models require maintaining hidden states across large regions and suffer from high latency due to strict serial decoding; (3) pure channel autoregression (Minnen & Singh) discards spatial local correlations, while pure spatial autoregression introduces decoding bottlenecks.

Key Challenge: The trade-off between modeling precision and computational efficiency — more expressive models better estimate the probability distribution of latent representations, but at the cost of escalating inference overhead and parameter counts.

Goal: Achieve state-of-the-art rate-distortion efficiency using a purely convolutional architecture, without relying on Transformers or recurrent components, while maintaining manageable parameter counts and inference speed.

Key Insight: Unify four complementary probabilistic modeling components — hierarchical hyperprior, masked spatial autoregression, channel-conditional modeling, and SE excitation — into a single VAE framework, where each component serves a distinct role rather than replacing the others.

Core Idea: Rather than pursuing larger and deeper models, the paper proposes more refined modeling of global, spatial, and channel-wise dependencies within a convolutional framework.

Method¶

Overall Architecture¶

ARCHE is built on a VAE framework: the analysis transform \(g_a\) maps the input to a latent representation \(y\), and the synthesis transform \(g_s\) reconstructs the image from the quantized representation \(\hat{y}\). The key contribution lies in the hierarchical design of the entropy model: the hyperprior provides global statistics → Masked PixelCNN context refines local probabilities → channel conditioning captures cross-channel dependencies → SE excitation adaptively weights channels → LRP corrects quantization error. The latent representation \(y\) is split into 10 slices along the channel dimension for sequential decoding, each with dedicated conditional transforms and LRP submodules.

Key Designs¶

Autoregressive Hyperprior + Masked PixelCNN Context
The hyper-analysis transform \(h_a\) maps \(y\) to side information \(z\), which is quantized and transmitted; the hyper-synthesis transform \(h_s\) reconstructs conditional prior parameters (mean \(\mu\) and scale \(\sigma\)) from \(\hat{z}\)
The spatial autoregressive prior uses Masked PixelCNN to model \(p(\hat{y}_i|\hat{y}_{<i}, \hat{z})\) in raster-scan order; Type A masks exclude the center and subsequent positions to ensure causality, while Type B masks include the center position
Stacked masked convolutional layers with sigmoid nonlinearities expand the receptive field, achieving more stable training and partially parallelizable inference compared to ConvLSTM
Hyperprior and context features are concatenated and passed through a parameter network (pointwise convolutions + small kernels + nonlinear activations) to produce the final Gaussian parameters
Channel Conditioning + SE Excitation
When decoding channel \(c\), features from the preceding \(c-1\) channels are used via lightweight convolutions to model the joint probability \(p(\hat{y}_{i,c}|\hat{y}_{<i,c}, \hat{y}_{<c}, \hat{z})\), extending the dependency space from purely spatial to spatial + channel
Squeeze-and-Excitation blocks are embedded within slice transforms: global average pooling produces channel descriptors → FC (reduction ratio 16) → ReLU → FC → sigmoid gating, adaptively amplifying informative channels and suppressing redundant ones
Cross-channel dependencies are typically smoother than spatial dependencies, allowing the channel-conditional module to remain lightweight without sacrificing performance
Latent Residual Prediction (LRP)
A correction is predicted for each quantized slice: \(\hat{y}'_m = \hat{y}_m + \lambda_{LRP} \cdot \text{softsign}(r_m)\)
Softsign replaces tanh to provide smoother gradients and bounded outputs; \(\lambda_{LRP}\) is a learnable scaling factor
LRP explicitly compensates for quantization noise that the hyperprior and context model cannot fully eliminate

Loss & Training¶

\(L = R + \lambda D\), where \(R\) is the cross-entropy rate (including the prior contribution from \(z\) and the conditional contribution from \(y|z\)), and \(D\) is MSE. Models are trained on the CLIC dataset with random \(256 \times 256\) crops normalized to \([0,1]\). Eight \(\lambda\) values \(\in \{0.001, 0.005, 0.007, 0.01, 0.03, 0.05, 0.07, 0.1\}\) cover different rate points. Optimizer: Adam with \(lr=10^{-4}\), 400 epochs, batch size 8. During training, quantization is approximated with additive uniform noise to maintain gradient flow. Latent depth: 320; 10 slices; hyperprior depth: 192; SE reduction ratio: 16. Implemented with TensorFlow 2.11 + TFC on an RTX 3080.

Key Experimental Results¶

Main Results¶

Method	BD-Rate vs Ballé (Kodak)	BD-Rate vs VVC (Kodak)	Params	Decode Time
Minnen et al.	-8.00%	+90.61%	95.8M	591ms
Minnen & Singh	-16.28%	+63.55%	121.7M	249ms
WeConvene	-6.92%	+92.47%	—	—
Iliopoulou et al.	-24.22%	+30.19%	124.3M	265ms
ARCHE	-48.01%	-5.61%	95.4M	222ms

On the Tecnick dataset: ARCHE achieves -44.89% vs. Ballé and -10.28% vs. VVC Intra, with consistent trends.

Ablation Study¶

Variant	Effect
Remove all AR components	Degrades to pure hyperprior; largest performance drop
Remove Masked Context Model	Significant degradation; spatial context is critical for local probability estimation
Remove SE	Moderate drop at low bitrates; channel weighting is important for preserving fine-grained structure
Slices 2→10	BD-Rate gain increases from ~5% to >11%; marginal returns beyond 10
GMM vs. single Gaussian	No significant improvement; conditional modeling already captures latent statistics sufficiently
Checkerboard vs. PixelCNN	58% faster training but worse rate-distortion (especially at low bitrates); inference is actually 15% slower

Key Findings¶

Component contributions are complementary rather than redundant; removing all simultaneously causes cumulative performance degradation
10 slices is the optimal balance: further splitting yields diminishing returns while computational cost grows linearly
ARCHE consistently outperforms VVC Intra on both Kodak and Tecnick datasets

Highlights & Insights¶

A purely convolutional architecture surpasses VVC Intra with better parameter efficiency and speed than most learned methods, providing strong evidence that carefully designed CNNs remain competitive
Each component collaborates within a unified probabilistic framework (confirmed by ablations), embodying a design philosophy of "specialized modeling" over "single complex modules"
Visual comparisons at low bitrates show sharper edges, more natural color transitions, and superior texture detail retention compared to VVC
Replacing LSTM context with Masked PixelCNN simultaneously improves training stability and inference speed

Limitations & Future Work¶

A 222ms decoding time remains insufficient for real-time video; block-wise semi-parallel decoding strategies could be explored
Optimization is limited to MSE; incorporating perceptual metrics (LPIPS/DISTS, etc.) could further improve visual fidelity
Task-oriented compression (e.g., using compressed representations directly for classification or segmentation) is not explored
Scalability and memory efficiency on higher-resolution images are not verified
Evaluation is limited to natural image datasets; generalization to medical imaging, remote sensing, and other domains remains unknown

vs. Iliopoulou et al. [2025]: ARCHE replaces LSTM context with Masked PixelCNN and adds SE excitation, achieving an additional ~24pp BD-Rate reduction, 29M fewer parameters, and 43ms faster decoding
vs. Minnen et al. [2018]: Building on joint AR + hyperprior, ARCHE adds channel conditioning, SE, and LRP, reducing BD-Rate by an additional ~40pp while cutting decoding time from 591ms to 222ms
vs. WeConvene [ECCV24]: The wavelet-domain approach performs more weakly (-6.92% vs. Ballé); ARCHE's spatial-domain joint modeling proves more effective
Design Philosophy: The principle of "better dependency modeling over larger models" and the complementary multi-level prior combination (global/spatial/channel) are transferable to other probabilistic modeling tasks

Rating¶

Novelty: ⭐⭐⭐ — Individual components have prior precedents; the contribution lies in careful integration and engineering optimization
Experimental Thoroughness: ⭐⭐⭐⭐⭐ — Dual datasets + 6 baselines + complete ablations + visual comparisons + computational analysis + appendix variant analysis
Writing Quality: ⭐⭐⭐⭐ — Detailed method derivation, rich tables and figures, transparent appendix
Value: ⭐⭐⭐ — Demonstrates that carefully designed CNN-based compression remains competitive; practically valuable for real-world deployment