Joint Degradation-Aware Arbitrary-Scale Super-Resolution for Variable-Rate Extreme Image Compression

Xinning Chai; Zhengxue Cheng; Xin Li; Rong Xie; Li Song

cs.CVMar 18, 2026

Joint Degradation-Aware Arbitrary-Scale Super-Resolution for Variable-Rate Extreme Image Compression

Authors: Xinning Chai, Zhengxue Cheng, Xin Li, Rong Xie, Li Song

Organizations: School of Information Science and Electronic Engineering, Shanghai Jiao Tong University · Department of Electronic Engineer and Information Science, University of Science and Technology of China · School of Information Science and Electronic Engineering, Shanghai Jiao Tong University and the MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University, China

Abstract

Recent diffusion-based extreme image compression methods have demonstrated remarkable performance at ultra-low bitrates. However, most approaches require training separate diffusion models for each target bitrate, resulting in substantial computational overhead and hindering practical deployment. Meanwhile, recent studies have shown that joint super-resolution can serve as an effective approach for enhancing low-bitrate reconstruction. However, when moving toward ultra-low bitrate regimes, these methods struggle due to severe information loss, and their reliance on fixed super-resolution scales prevents flexible adaptation across diverse bitrates. To address these limitations, we propose ASSR-EIC, a novel image compression framework that leverages arbitrary-scale super-resolution (ASSR) to support variable-rate extreme image compression (EIC). An arbitrary-scale downsampling module is introduced at the encoder side to provide controllable rate reduction, while a diffusion-based, joint degradation-aware ASSR decoder enables rate-adaptive reconstruction within a single model. We exploit the compression- and rescaling-aware diffusion prior to guide the reconstruction, yielding high fidelity and high realism restoration across diverse compression and rescaling settings. Specifically, we design a global compression-rescaling adaptor that offers holistic guidance for rate adaptation, and a local compression-rescaling modulator that dynamically balances generative and fidelity-oriented behaviors to achieve fine-grained, bitrate-adaptive detail restoration. To further enhance reconstruction quality, we introduce a dual semantic-enhanced design. Extensive experiments demonstrate that ASSR-EIC delivers state-of-the-art performance in extreme image compression while simultaneously supporting flexible bitrate control and adaptive rate-dependent reconstruction.

Figures & tables

Fig. 1: Framework overview of the proposed ASSR-EIC and its rate–distortion performance. Left: Framework overview, which integrates arbitrary-scale downsampling and diffusion-based joint degradation-aware ASSR reconstruction to achieve high-quality, bitrate-adaptive extreme image compression. Right: R–D (CLIPScore) curves on MSCOCO demonstrate that ASSR-EIC outperforms state-of-the-art extreme image compression methods PerCo [ 1 ] and DiffEIC [ 2 ] , while supporting variable bitrates, illustrated using MS-ILLM (“quality 1”) as the anchor codec with different rescaling factors $s$ .

Fig. 2: Structure overview of the proposed ASSR-EIC. We insert an arbitrary-scale downsampler to achieve controllable bitrate reduction at the encoder end, and propose a joint degradation-aware ASSR reconstruction decoder to restore the high-quality HR image with rate adaptation. The joint degradation-aware ASSR reconstruction decoder consists of a backbone, a fidelity module, a global compression-rescaling adaptor, and a local compression-rescaling modulator. The layers of the Backbone and the Fidelity Module are arranged from left to right to indicate the forward feature propagation order, with arrows omitted for clarity. $s,\chi_{CT},\chi_{QP}$ denote the rescaling factor, codec type, and codec quality parameter, respectively.

Codec	Quality Parameter	bpp (Orig.)	bpp ( $s{=}1.2$ )	bpp ( $s{=}1.5$ )	bpp ( $s{=}2$ )
VVC	QP=47	0.124	0.080 (1.55 $\times$ )	0.059 (2.10 $\times$ )	0.040 (3.10 $\times$ )
	QP=52	0.057	0.042 (1.36 $\times$ )	0.033 (1.73 $\times$ )	0.024 (2.38 $\times$ )
MS-ILLM	”quality_2”	0.091	0.069 (1.32 $\times$ )	0.054 (1.69 $\times$ )	0.029 (3.14 $\times$ )
	”quality_1”	0.053	0.042 (1.26 $\times$ )	0.033 (1.61 $\times$ )	0.019 (2.79 $\times$ )
Avg. ∗		–	$\approx$ 1.37 $\times$ ( $\approx 0.95\,s^{2}$ )	$\approx$ 1.78 $\times$ ( $\approx 0.79\,s^{2}$ )	$\approx$ 2.85 $\times$ ( $\approx 0.71\,s^{2}$ )

TABLE I: Impact of Spatial Downsampling on Bitrate for Representative Anchor Codecs on MSCOCO. The value in parentheses denotes the ratio

\mathrm{bpp}_{\mathrm{orig}}/\mathrm{bpp}_{\mathrm{down}}

, indicating how many times the bitrate is reduced after spatial downsampling.

Fig. 3: Details of the encoding embedding generation. We encode the float-type compression parameters and the rescaling factor into feature embeddings.

Fig. 4: Detailed structure of the global compression-rescaling adaptor.

Fig. 5: Detailed structure of the local compression-rescaling modulator.

Fig. 6: Evaluation of ASSR-EIC and competing compression codecs on the Kodak, MSCOCO, and CLIC2020 datasets. FID is not reported on Kodak because its limited dataset size is insufficient for reliable FID estimation.

Fig. 7: Visual comparison of our ASSR-EIC against representative state-of-the-art image compression codecs.

Fig. 8: Rate–distortion performance improvements of our ASSR-EIC over the basic VVC and MS-ILLM anchors on the Kodak, MSCOCO, and CLIC2020. FID is not reported on Kodak because its limited dataset size is insufficient for reliable FID estimation.

Fig. 9: Subjective visual improvement results of ASSR-EIC against the basic anchor model under different rescaling factors $s$ , showing its ability to maintain high perceptual realism even as bitrate decreases.

Fig. 10: Impact of the rescaling factor $s$ on the bitrate and reconstruction quality of ASSR-EIC, evaluated using VVC and MS-ILLM as anchor codecs. and represent the results of the VVC (QP=52) and MS-ILLM (“quality_1”) anchor models, respectively.

Anchor	ASSR-EIC (Anchor)
Anchor	FID-BPP↓	$d_{clip}$ -BPP↓	DISTS-BPP↓
VVC	-93.26	-83.70	-83.27
MS-ILLM	-30.96	-31.90	-14.44

TABLE II: BD-Rate (%) improvements achieved by our ASSR-EIC over the basic VVC and MS-ILLM anchor codecs on MSCOCO.

Anchor	SR method	FID-BPP $\downarrow$	$d_{clip}$ -BPP $\downarrow$	DISTS-BPP $\downarrow$
VVC	– (anchor only)	0%	0%	0%
	RealESRGAN [ 88 ]	7.74%	-5.34%	-9.81%
	PromptCIR [ 89 ]	4.23%	-0.68%	9.96%
	DA-CLIP [ 90 ]	-0.55%	0.67%	0.22%
	SeeSR [ 62 ]	-72.70%	-55.49%	-73.85%
	ASSR-EIC (Ours)	-93.26%	-83.70%	-83.70%

TABLE III: Comparison with alternative super-resolution (SR) methods as the post-processing module after downsampling. We report BD-Rate (%) on MS-COCO; lower is better.

Anchor	ASSR-EIC (Anchor)
Anchor	FID-BPP↓	$d_{clip}$ -BPP↓	DISTS-BPP↓
HEVC	-87.97	-76.14	-87.29
ELIC	-89.20	-82.66	-86.92

TABLE IV: Generalization performance of our ASSR-EIC framework when applied to unseen codecs. BD-Rate (%) results between our approach and anchor methods.

Method	FID-BPP↓	$d_{clip}$ -BPP↓	DISTS-BPP↓
w/o Textual Guidance	6.48	2.29	3.07
w/o Semantic Attention	6.69	8.78	8.58
ASSR-EIC	0	0	0

TABLE V: Ablation study of the dual semantic-enhanced design, showing the individual contributions of textual guidance and the soft semantic prompt, demonstrated by BD-Rate(%).

Fig. 11: Ablation results of key components in ASSR-EIC, including the global compression–rescaling adaptor and the local compression–rescaling modulator, showing their impacts on FID and DISTS across different bitrates.

Type	Method	Training Steps	Encoding Speed (s)	Decoding Speed (s)	Inference Step	MACs (G)
Non-Diffusion-based	VVC	-	18.89±7.42	0.62±0.06	1	-
	MS-ILLM	N×2M	0.08±0.02	0.06±0.02	1	200
	ELIC	N×2M	0.20±0.01	0.12±0.04	1	155
Diffusion-based	Text-Sketch	-	137.28±2.55	9.45±0.53	20	303053
	PerCo	N×280k	0.42±0.04	1.52±0.01	20	15934
	DiffEIC	N×500k	0.33±0.01	3.05±0.02	50	36491

TABLE VI: Complexity comparison between our ASSR-EIC framework and existing image compression methods, including training cost, encoding time, decoding time, and MACs. N denotes the number of models required to support different bitrates. The suffix “S” in diffusion-based methods (e.g., ASSR-EIC-S20 and ASSR-EIC-S50) denotes the number of diffusion inference steps.

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

Jun 1, 2026

Exploiting Semantic and Pixel Representations for Ultra-Low Bitrate Image Compression

Most existing extreme compression methods fail to achieve an optimal rate-distortion-perception trade-off, as they typically prioritize perceptual fidelity and visual realism over pixel-level accuracy. Consequently, the resulting reconstructions often deviate noticeably from the originals. Ultra-low bitrate image compression is therefore crucial-not only for producing extremely compact representations but also for ensuring that reconstructed images remain semantically coherent and faithful to the source at the pixel level. To this end, we propose SPRDiff, a diffusion-based compression method that fully leverages both semantic and pixel representations, thereby enhancing reconstruction fidelity under ultra-low bitrate constraints. Specifically, we develop a triple-encoder architecture that utilizes high-fidelity features from the pretrained distortion-oriented and semantic-oriented encoders to compensate for the limited representations extracted by the frozen VAE encoder, thereby improving latent compression and entropy modeling. To further enhance the reconstruction fidelity of diffusion models, we introduce a distortion-aware reconstruction module with dual feature extraction. This module not only generates a coarse reconstruction that preserves the main structures, but also provides practical and accurate semantic- and pixel-level conditional signals to guide the diffusion model. Extensive experiments on benchmark datasets demonstrate that our method outperforms state-of-the-art approaches in the rate-distortion-perception tradeoff at extremely low bitrates (below 0.03 bpp), effectively preserving both perceptual quality and pixel-wise fidelity in the reconstructed images. We will release the source code and trained models at https://github.com/cshw2021/SPRDiff.

Hao Wei, Yanhui Zhou, Chenyang Ge +2

National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University, Xi’an 710049, China · School of Information and Telecommunication, Xi’an Jiaotong University, Xi’an 710049, China · Department of Computer Science and Software Engineering, The University of Western Australia, Perth, Crawley, WA 6009, Australia

Learned Image Compression Diffusion Models

Ask PDF

Jul 15, 2026

MixCompress: Mixture of Experts for Variable Rate Learned Image Compression

Learned image compression (LIC) is bottlenecked by the need to store independent models for each rate-distortion operating point. Existing variable bit-rate (VBR) methods aim to reduce this overhead via dense parameter modulation, but forcing a shared backbone to approximate divergent mappings causes severe feature entanglement. Specifically, low-rate smoothing gradients inherently conflict with the preservation of high-frequency textural details, leading to sub-optimal performance. To resolve this, we propose MixCompress, a unified VBR framework based on sparse structural specialization. While sparsely gated Mixture-of-Experts (MoE) routing successfully mitigates gradient conflict, it operates on a fixed computational budget. To address the increased representational demands of higher bit-rates we introduce a Mixture-of-Depths (MoD) extension to dynamically scale model capacity. Combined with Conditional Auxiliary Transforms (CAT) for dynamic sub-band energy modulation, our hierarchical framework effectively dynamically scales capacity. Extensive evaluations demonstrate that MixCompress not only matches individually optimized single-rate baselines but can even surpass them, establishing a new Pareto frontier for computationally efficient image coding.

Calvin-Khang Ta, Praneet Singh, Tong Shao +1

Dolby Laboratories, USA

Learned Image Compression Rate-Distortion Optimization Mixture-of-Experts Models

Ask PDF

Sep 30, 2026

ResARC: Residual-Aware AutoRegressive Coding for Ultra-Low Bitrate Image Compression

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

Learned Image Compression Residual Learning Generative Image Compression

Ask PDF

Abstract

Learned Image Compression

Explore similar work

Exploiting Semantic and Pixel Representations for Ultra-Low Bitrate Image Compression

MixCompress: Mixture of Experts for Variable Rate Learned Image Compression

ResARC: Residual-Aware AutoRegressive Coding for Ultra-Low Bitrate Image Compression