Adaptive Learned Image Compression with Graph Neural Networks

Yunuo Chen; Bing He; Zezheng Lyu; Hongwei Hu; Qunshan Gu; Yuan Tian; Guo Lu

cs.CVMar 26, 2026

Adaptive Learned Image Compression with Graph Neural Networks

Authors: Yunuo Chen, Bing He, Zezheng Lyu, Hongwei Hu, Qunshan Gu, Yuan Tian, Guo Lu

Organizations: Shanghai Jiao Tong University · Massachusetts Institute of Technology · Alibaba Group · Shanghai AI Laboratory

Abstract

Efficient image compression relies on modeling both local and global redundancy. Most state-of-the-art (SOTA) learned image compression (LIC) methods are based on CNNs or Transformers, which are inherently rigid. Standard CNN kernels and window-based attention mechanisms impose fixed receptive fields and static connectivity patterns, which potentially couple non-redundant pixels simply due to their proximity in Euclidean space. This rigidity limits the model's ability to adaptively capture spatially varying redundancy across the image, particularly at the global level. To overcome these limitations, we propose a content-adaptive image compression framework based on Graph Neural Networks (GNNs). Specifically, our approach constructs dual-scale graphs that enable flexible, data-driven receptive fields. Furthermore, we introduce adaptive connectivity by dynamically adjusting the number of neighbors for each node based on local content complexity. These innovations empower our Graph-based Learned Image Compression (GLIC) model to effectively model diverse redundancy patterns across images, leading to more efficient and adaptive compression. Experiments demonstrate that GLIC achieves state-of-the-art performance, achieving BD-rate reductions of 19.29%, 21.69%, and 18.71% relative to VTM-9.1 on Kodak, Tecnick, and CLIC, respectively. Code will be released at https://github.com/UnoC-727/GLIC.

Figures & tables

Figure 2: Overview of our method. (a) Architecture of the proposed GLIC codec. Channel widths are $C_{1},C_{2},C_{3},C_{4}$ , and the numbers of non-linear transform blocks are $L_{1},L_{2},L_{3}$ . (b) Graph-based Feature Aggregation Block used as advanced non-linear transforms. (c) Lightweight Conv Block for early feature extraction and reduced complexity.

Method	Candidates	Selection & Degree	Receptive Field	Complexity
Canonical operators
Dynamic CNN	$k{\times}k$ window	keep-all; degree $=k^{2}$	bounded by window	$\mathcal{O}(1)\;+\;\mathcal{O}(Nk^{2})$
Deformable CNN	offset-enlarged $k{\times}k$ window	keep-all; degree $=k^{2}$	offset-enlarged local window	$\mathcal{O}(Nk)\;+\;\mathcal{O}(Nk)$
Window Attention	$L{\times}L$ window	keep-all; degree $=L^{2}$	bounded by window	$\mathcal{O}(1)\;+\;\mathcal{O}(NL^{2})$
Transformer (global)	all nodes	keep-all; degree $=N$	full	$\mathcal{O}(1)\;+\;\mathcal{O}(N^{2})$
GNN-based LIC

Table 1: Connectivity-oriented comparison. Notation:

N

denotes the total number of nodes/pixels;

k

the side length of the convolution kernel;

L

the side length of the attention/GNN window; and

K

the average neighborhood size (connectivity and in-degrees for GNNs). Computational cost is reported as construction + aggregation . “Construction” counts candidate search and similarity calculation.

Figure 5: PSNR and MS-SSIM R-D curves on the Kodak Dataset.

Method	Complexity					BD‑rate (%) $\downarrow$
Method	Params (M)	FLOPs (T)	Enc‑Lat. (s)	Dec‑Lat. (s)	Peak‑Mem. (G)	Kodak	Tecnick	CLIC
VTM‑9.1 [ 7 ]	-	-	-	-	-	0.00	0.00	0.00
ELIC (CVPR’22) [ 26 ]	33.29	1.74	0.583	0.335	1.50	-5.95	-7.68	-1.20
MLIC++ (ICML’23W) [ 29 ]	116.48	2.64	0.508	0.547	2.08	-15.14	-17.23	-14.41
TCM-L (CVPR’23) [ 47 ]	75.89	3.74	0.647	0.542	7.73	-13.42	-10.93	-9.10
FTIC (ICLR’24) [ 38 ]	69.78	2.38	$>$ 10	$>$ 10	4.90	-14.83	-14.39	-10.70

Table 2: Quantitative Comparisons of leading LIC methods. BD‑rates are evaluated against VTM‑9.1. FLOPs, peak memory and decoding time are measured on 2K‑resolution images.

Panel of Figure 6 — Figure 6: (Top) Averaged Effective Receptive Fields of state-of-the-art LIC models across the 24 Kodak images. (Bottom) Single-image Effective Receptive Field for Kodak image 23, demonstrating the adaptivity of our GLIC model. More results are in the supplementary.

Figure 7: Spatial redundancy versus RMS-G scores. Image023 from the Kodak dataset (left) and the RMS-G maps for the second- and third-stage graphs in GLIC (right). Brighter regions have higher RMS-G scores, indicating lower spatial redundancy and larger in-degrees.

Figure 8: Visual Comparisons with other methods. More results are in the supplementary.

Variant	Kodak	CLIC	Tecnick
w/o Global (Local only)	-16.98	-15.71	-18.89
w/o Local (Global only)	-15.62	-15.02	-17.51
Reverse (Local+Global)	-18.77	-18.21	-20.79
Global+Local (Ours)	-19.29	-18.71	-21.69

Table 3: Ablations on dual-scale graph flexibility.

Scoring Strategy	Channel Pooling	Kodak	CLIC	Tecnick
None	None	-16.97	-16.21	-18.21
Local Entropy	RMS	-17.05	-17.01	-18.97
Rescaling Residual	RMS	-17.67	-17.03	-19.68
Rescaling Residual	Mean	-18.23	-17.82	-20.39
Sobel Gradient	Mean	-18.02	-17.42	-20.62
Sobel Gradient	RMS	-19.29	-18.71	-21.69

Table 4: Ablations on complexity scoring and channel pooling.

Learned Image Compression

Learned image compression uses neural encoders, quantization, and entropy models to reduce image bitrates while balancing fidelity, perceptual quality, and task utility.

See all 120 papers

Efficient Learned Image Compression without Entropy Coding
May 22, 2026Hao Cao, Wenqi Guo, Zhijin Qin +1Learned Image Compression Vector Quantization
Department of Electronic Engineering, Tsinghua University · Department of Automation, Tsinghua University · State Key Laboratory of Space Network and Communications +1
Variable-Rate Deep Image Compression based on Low-Rank Adaptation by Progressive Learning
Jun 15, 2026Xing-Yu Xu, Chen-Hsiu Huang, Ja-Ling WuLearned Image Compression Low-Rank Adaptation
Communications and Multimedia Lab National Taiwan University Taipei, Taiwan
MixCompress: Mixture of Experts for Variable Rate Learned Image Compression
Jul 15, 2026Calvin-Khang Ta, Praneet Singh, Tong Shao +1Learned Image Compression Rate-Distortion Optimization
Dolby Laboratories, USA

May 22, 2026J/K move · Enter open · S save

Efficient Learned Image Compression without Entropy Coding

Hao Cao, Wenqi Guo, Zhijin Qin, Jungong Han

Department of Electronic Engineering, Tsinghua University · Department of Automation, Tsinghua University · State Key Laboratory of Space Network and Communications+1

Learned Image Compression Vector Quantization

Open paper AskPDF arXiv

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\times

faster encoding and

5\times

faster decoding.

Abstract

Learned Image Compression

Explore similar work

Efficient Learned Image Compression without Entropy Coding

Variable-Rate Deep Image Compression based on Low-Rank Adaptation by Progressive Learning

MixCompress: Mixture of Experts for Variable Rate Learned Image Compression