DACESR: Degradation-Aware Conditional Embedding for Real-World Image Super-Resolution

Xiaoyan Lei; Wenlong Zhang; Biao Luo; Hui Liang; Weifeng Cao; Qiuting Lin

cs.CVFeb 27, 2026

DACESR: Degradation-Aware Conditional Embedding for Real-World Image Super-Resolution

Authors: Xiaoyan Lei, Wenlong Zhang, Biao Luo, Hui Liang, Weifeng Cao, Qiuting Lin

Organizations: School of Electrical and Information Engineering, Zhengzhou University of Light Industry, Zhengzhou 450002, Henan, China. · Shanghai Artificial Intelligence Laboratory, Shanghai 200233, China. · School of Automation, Central South University, Changsha 410083, China. · Intelligent Manufacturing Engineering at Machinery Technology Development Co.,Ltd., China Academy of Machinery Science and Technology Group Co.,Ltd., Beijing 100801, China.

Abstract

Multimodal large models have shown excellent ability in addressing image super-resolution in real-world scenarios by leveraging language class as condition information, yet their abilities in degraded images remain limited. In this paper, we first revisit the capabilities of the Recognize Anything Model (RAM) for degraded images by calculating text similarity. We find that directly using contrastive learning to fine-tune RAM in the degraded space is difficult to achieve acceptable results. To address this issue, we employ a degradation selection strategy to propose a Real Embedding Extractor (REE), which achieves significant recognition performance gain on degraded image content through contrastive learning. Furthermore, we use a Conditional Feature Modulator (CFM) to incorporate the high-level information of REE for a powerful Mamba-based network, which can leverage effective pixel information to restore image textures and produce visually pleasing results. Extensive experiments demonstrate that the REE can effectively help image super-resolution networks balance fidelity and perceptual quality, highlighting the great potential of Mamba in real-world applications. The source code of this work will be made publicly available at: https://github.com/nathan66666/DACESR.git

Figures & tables

Fig. 1 : The tag representations of RAM on clean images and images with varying levels of degradation.“Similarity” refers to the Jaccard similarity (Eq. ( 1 )).

Panel of Figure 2 — Fig. 2 : Comparison of text output accuracy across RAM under different types and intensities of degradation. (a) Blur: The x-axis represents isotropic Gaussian blur, where larger values indicate stronger blurring. (b) JPEG: The x-axis denotes JPEG compression levels, with lower values indicating higher compression. (c) Noise: The x-axis represents the intensity of additive Gaussian noise, where higher values correspond to increased noise levels. In (d), the classification is based on the text similarity values of RAM for different degraded outputs, which are evenly divided into four categories in descending order. Each category contains multiple types/levels of degradation.

Fig. 4 : The training pipeline of the Real Embedding Extractor (REE).

Datasets	Metric	ESRGAN	BSRGAN	Real-ESRGAN	SwinIR	DASR	HAT	KDSR	DCLS	CDFormer	DACESR
Datasets	Metric	[ 13 ]	[ 14 ]	[ 15 ]	[ 10 ]	[ 16 ]	[ 11 ]	[ 50 ]	[ 51 ]	[ 52 ]	(Ours)
Bicubic	$\mathrm{PSNR}~(\uparrow)$	28.17	27.32	26.65	27.21	28.55	27.44	27.48	30.73	30.87	28.34
Bicubic	$\mathrm{LPIPS}~(\downarrow)$	0.115	0.236	0.228	0.214	0.170	0.203	0.295	0.291	0.286	0.139
Level-1	$\mathrm{PSNR}~(\uparrow)$	21.16	26.78	26.17	26.45	27.84	26.85	26.69	27.32	26.90	28.16
Level-1	$\mathrm{LPIPS}~(\downarrow)$	0.473	0.241	0.231	0.216	0.171	0.202	0.307	0.368	0.363	0.142
Level-2	$\mathrm{PSNR}~(\uparrow)$	22.77	26.75	26.16	26.39	27.58	26.81	26.76	27.24	26.92	27.68

TABLE I : Quantitative comparison with state-of-the-art methods for real-world image super-resolution on benchmark datasets. For the compared methods, we employ their officially released pre-trained models. The best and second best results are highlighted in red and blue respectively.

Fig. 5 : The LAM results of different model architectures across various types of degradation. LAM attribution indicates the significance of each pixel in the input LR image during the reconstruction process of the patch highlighted by a box. The Diffusion Index (DI) denotes the extent of pixel involvement. A higher DI indicates a broader range of utilized pixels.

Fig. 6 : Real-world image super-resolution results on SR $\times$ 4.

Network	Params(M)	FLOPS(G)	Time(ms)	Bicubic	Level-I	Level-II	Level-III
Network	Params(M)	FLOPS(G)	Time(ms)	$\mathrm{PSNR}~(\uparrow)/\mathrm{LPIPS}~(\downarrow)$	$\mathrm{PSNR}~(\uparrow)/\mathrm{LPIPS}~(\downarrow)$	$\mathrm{PSNR}~(\uparrow)/\mathrm{LPIPS}~(\downarrow)$	$\mathrm{PSNR}~(\uparrow)/\mathrm{LPIPS}~(\downarrow)$
SRResNet	1.52	166	113	28.05/0.1747	27.60/0.1772	27.34/0.2228	23.71/0.4419
EDSR	1.52	130	105	28.26/0.1807	27.79/0.1834	27.53/0.2284	23.87/0.4351
SwinIR	11.72	539	1719	28.28/0.1488	27.78/0.1531	27.45/0.1854	23.60/0.3869
RRDB	16.70	1176	460	27.92/0.1473	27.84/0.1569	27.29/0.1886	23.54/0.3847
DASR	8.07	184	142	28.55 /0.1696	27.84/0.1707	27.58/0.2126	23.93/0.4144

TABLE II : Quantitative comparison of different backbone networks for real-world image super-resolution, all retrained on the same degradation space. FLOPS are measured with a

256\times 256

input, and all data are obtained using an NVIDIA V100 GPU. “Mamba-based” is a Mamba-based method, derived from [ 30 ] . The best results are highlighted in bold.

Method	DrealSR [ 58 ]	RealSR [ 53 ]
Method	$\mathrm{PSNR}~(\uparrow)/\mathrm{LPIPS}~(\downarrow)$	$\mathrm{PSNR}~(\uparrow)/\mathrm{LPIPS}~(\downarrow)$
StableSR [ 41 ]	28.13/0.3315	24.70/0.3018
PASD [ 20 ]	27.00/0.3931	24.29/0.3435
DiffBIR [ 21 ]	26.76/0.4599	24.77/0.3658
SeeSR [ 19 ]	28.17/0.3189	25.18/0.3009
Ours	29.74 / 0.2803	27.31 / 0.2911

TABLE III : Quantitative comparison with diffusion-based methods. The data for the methods we compared is sourced from [ 19 ] . The best results are highlighted in bold.

Backbone	Condition	Bicubic	Level-I	Level-II	Level-III
Backbone	Condition	$\mathrm{PSNR}~(\uparrow)/\mathrm{LPIPS}~(\downarrow)$	$\mathrm{PSNR}~(\uparrow)/\mathrm{LPIPS}~(\downarrow)$	$\mathrm{PSNR}~(\uparrow)/\mathrm{LPIPS}~(\downarrow)$	$\mathrm{PSNR}~(\uparrow)/\mathrm{LPIPS}~(\downarrow)$
Mamba-based	None	28.05 /0.1811	27.57 /0.1840	27.39 /0.2209	23.80 / 0.3938
Mamba-based	RE	27.79/ 0.1546	27.36/ 0.1676	27.08/ 0.1989	23.73/0.3975
Mamba-based	LE	27.56/0.1945	27.23/0.1903	27.07/0.2227	23.67/0.4078

TABLE IV : Quantitative Comparison of Different Conditional Information Prompts. “Mamba-based” refers to the backbone composed of RSSB [ 30 ] . “RE” represents the representation embedding of the low-resolution image, and “LE” represents the logits embedding of the low-resolution image.

Fig. 7 : Visual comparison of results with and without conditional information under different degradations using the same backbone network.

Backbone	Condition	Bicubic	Level-I	Level-II	Level-III
Backbone	Network	$\mathrm{PSNR}~(\uparrow)/\mathrm{LPIPS}~(\downarrow)$	$\mathrm{PSNR}~(\uparrow)/\mathrm{LPIPS}~(\downarrow)$	$\mathrm{PSNR}~(\uparrow)/\mathrm{LPIPS}~(\downarrow)$	$\mathrm{PSNR}~(\uparrow)/\mathrm{LPIPS}~(\downarrow)$
Mamba-based	RAM	27.72/0.1761	27.27/0.1770	27.08/0.2067	23.69/ 0.3860
Mamba-based	DAPE	27.45/0.1720	27.25/0.1740	26.96/0.2062	23.68/0.3931
Mamba-based	REE	27.79 / 0.1546	27.36 / 0.1676	27.08 / 0.1989	23.73 /0.3975

TABLE V : Quantitative Comparison of Different Conditional Networks. “Mamba-based” refers to the RSSB [ 30 ] .

Fig. 8 : Using images with different degradations as input, the average activation maps of the intermediate layers of each network.

$\times$	$+$	Bicubic	Level-I	Level-II	Level-III
$\times$	$+$	$\mathrm{PSNR}/\mathrm{LPIPS}$	$\mathrm{PSNR}/\mathrm{LPIPS}$	$\mathrm{PSNR}/\mathrm{LPIPS}$	$\mathrm{PSNR}/\mathrm{LPIPS}$
✔	✘	27.80 /0.1877	27.33/0.1858	27.10/0.2203	23.52/0.4011
✘	✔	27.71/0.1789	27.38 /0.1775	27.12 /0.2123	23.58/0.3982
✔	✔	27.79/ 0.1546	27.36/ 0.1676	27.08/ 0.1989	23.73 / 0.3975

TABLE VI : Quantitative Comparison of Conditioned Information Modulator Designs.

Image Super-Resolution

Image super-resolution reconstructs high-resolution images from lower-resolution observations, balancing detail, fidelity, and computational cost.

See all 248 papers

Super-Resolution in The Right Latent Space: A Frozen Vision-Foundation Substrate
Oct 3, 2026Wanzhou Lei, Cuifeng Shen, Yanjin He +6Image Super-Resolution
University of California, Berkeley · Alibaba Group · University of Michigan, Ann Arbor +2
Degradation-Aware and Structure-Preserving Diffusion for Real-World Image Super-Resolution
Apr 13, 2026Yang Ji, Zonghao Chen, Zhihao Xue +1Degradation-Aware Image Restoration Image Super-Resolution
KUNBYTE Hangzhou, Zhejiang Province, China · GOLDMYE Xuhui District, Shanghai, China
VICR: Visual In-Context Restoration for Real-World Image Super-Resolution
May 30, 2026Qichang Zhang, Hailong Wang, Baiang Li +4Diffusion Models for Image Restoration Image Super-Resolution
Faculty of Science and Technology, University of Macau · 2Nullmax · 3Hefei University of Technology +1

Oct 3, 2026J/K move · Enter open · S save

Super-Resolution in The Right Latent Space: A Frozen Vision-Foundation Substrate

Wanzhou Lei, Cuifeng Shen, Yanjin He, Maohua Li+5

University of California, Berkeley · Alibaba Group · University of Michigan, Ann Arbor+2

Image Super-Resolution

Open paper AskPDF arXiv

In an image latent space, the embeddings of high-resolution, natural, and sharp images form a manifold. Degradation of high-resolution images pushes their embeddings off this manifold. Real-world super-resolution (SR) then becomes the task of mapping the degraded embedding back onto this manifold --- not anywhere on the manifold, but to the point that preserves what the input still carries, both its semantics and pixel details. Every published method implements this mapping in a reconstruction-oriented latent space or pixel space. We claim these spaces are the wrong substrates for SR. Low-resolution and degraded images are embedded far from the manifold, making the mapping difficult and expensive. The lack of semantic information in these substrates also makes it difficult to navigate to the faithful point on the manifold, causing severe hallucination when degradation is heavy. Thus, restoring in a suitable latent space is crucial to the SR task. We show that the latent space of 23 fused layers of a frozen DINOv3-L is one such space that makes the SR task easier. Degraded images are embedded near the manifold. Moreover, this substrate contains a hierarchy of information, from pixel record to degradation robust semantics, guiding the model to find the faithful point on the manifold. On this substrate, a 415M decoder is trained under reconstruction and adversarial objectives to map the degraded embeddings back and decode to pixel space in one pass. The resulting model, RAESR, attains the best fidelity--perception trade-off among state-of-the-art adversarial and diffusion-based restorers on RealSR, DRealSR, LSDIR and DIV2K-Val, at 37 ms per 512 by 512 image on a single H20 GPU. Swapping the substrate for a VAE latent under an identical recipe loses on every metric.

Abstract

Image Super-Resolution

Explore similar work

Super-Resolution in The Right Latent Space: A Frozen Vision-Foundation Substrate

Degradation-Aware and Structure-Preserving Diffusion for Real-World Image Super-Resolution

VICR: Visual In-Context Restoration for Real-World Image Super-Resolution