Image super-resolution in the dark is fundamentally challenged by extreme spatial heterogeneity: severely underexposed and noisy regions demand aggressive generative enhancement to reconstruct missing details, while relatively well-exposed areas require conservative updates to prevent hallucinated textures. Standard diffusion and flow-matching models, however, rely on globally uniform integration trajectories, making them sub-optimal for such spatially varying degradations. In this paper, we propose DeLang-SR (Degradation-Language steered Super-Resolution), a one-step flow-matching framework that explicitly decouples the restoration direction and magnitude. Rather than treating low-light corruption as arbitrary latent features, DeLang-SR translates measurable physical statistics into structured degradation language, harnessing the semantic prior of pretrained text-to-image models. The degradation-language intent, complemented by image-specific visual tokens, steers the velocity field (direction) toward restoration-relevant manifolds. Simultaneously, a continuous spatial update-strength field acts as a locally varying Euler step-size field (magnitude), applying stronger integration steps in heavily degraded shadows while enforcing conservative updates in reliable regions. Experiments on paired and real dark images, together with ablations and diagnostic analyses, show that DeLang-SR provides an effective and interpretable one-step route to super-resolution in the dark. Code and pretrained models are available at https://github.com/moriyaya/DeLang-SR.
Figures & tables
Figure 1: Pipeline of DeLang-SR. Degradation language and visual tokens form Hd to steer the SD3 velocity direction, while the continuous descriptor and local evidence predict Mϕ(y,δ) to control spatial update magnitude. During inference, (δ,ℓd) are estimated by Vtest(y) without GT, and the restored image is obtained by a single direction-steered and magnitude-controlled flow update.
Figure 2: Visualization of the update-strength map Mϕ(y,δ) . Both ℓd and δ are estimated from y without GT. The map provides an interpretable spatial step-size field for one-step SD3 restoration.
Figure 3: Inference-time condition source analysis. The close overlap among generic, LQ-only, and oracle intents shows that DeLang-SR does not require oracle degradation annotations.
Method
Type
RELLISUR @ ×2
RELLISUR @ ×4
PSNR ↑
SSIM ↑
LPIPS ↓
DISTS ↓
ExpErr ↓
ColorErr ↓
PSNR ↑
SSIM ↑
LPIPS ↓
DISTS ↓
ExpErr ↓
ColorErr ↓
† RUAS → ‡ SwinIR
♣
17.160
0.6170
0.5811
0.3048
0.1024
0.3074
17.058
0.6624
0.6633
0.3756
0.0871
0.2611
† SCI → ‡ HAT
♣
14.963
0.4389
0.5968
0.3381
0.1644
0.5038
14.750
0.4523
0.7040
0.3591
0.1644
0.5057
† LLFormer → ‡ HAT
♣
21.221
0.7205
0.4617
0.2660
0.0549
0.1683
20.123
0.7182
0.5847
0.3639
0.0636
0.1951
SwinIR (ICCVW’21)
▲
18.389
0.6407
0.5866
0.2862
0.0666
0.2049
17.513
0.6637
0.6968
0.3922
0.0797
0.2445
Restormer (CVPR’22)
▲
21.219
0.7279
0.3901
0.2441
0.0593
0.1800
20.276
0.7207
0.5029
0.3179
0.0660
0.2025
Table 1: Paired quantitative comparison on RELLISUR . Pink , orange , and yellow mark the top three results. ♣ , ▲ , and ⧫ denote cascaded LLE+SR, general SR, and diffusion/degradation-based methods, respectively. † / ‡ indicate LLE-trained/normal-light SR-trained models.
Method
Type
RELLISUR @ ×2
RELLISUR @ ×4
NIQE ↓
MUSIQ ↑
MANIQA ↑
CLIP-IQA ↑
NIQE ↓
MUSIQ ↑
MANIQA ↑
CLIP-IQA ↑
HiT-SR (ECCV’24)
▲
7.527
26.2128
0.2743
0.2580
9.391
21.2099
0.2231
0.1851
PFT-SR (CVPR’25)
▲
8.298
31.3606
0.2217
0.2283
9.980
23.3443
0.2460
0.2844
CATANet (CVPR’25)
▲
8.015
35.5178
0.2003
0.1891
8.735
21.5868
0.2094
0.1966
DHGM (AAAI’26)
⧫
8.795
41.7819
0.2140
0.2050
9.929
25.3630
0.2297
0.2277
DOD (AAAI’26)
⧫
7.357
50.2245
0.2819
0.4224
8.610
29.0542
0.2589
0.4935
Table 2: No-reference perceptual quality comparison. Top-three results are highlighted.
Figure 4: Distribution of no-reference perceptual quality scores on DarkFace under the ×2 setting.
Figure 5: Visual comparison of all competing methods on RELLISUR (Top two:×2, Bottom two:×4). More RELLISUR visual comparisons under ×2 and ×4 settings are provided in Appx. G.6 .
Figure 6: Visual comparison on DarkFace under the ×2 setting.
Metric
HiT-SR
ASID
PFT-SR
CATANet
RGT
DHGM
DOD
DeLang-SR
Trainable Params. (M) ↓
0.772
0.298
19.624
0.477
10.051
10.75
958.684
4.48
FLOPs (G) ↓
18.290
3.965
106.359
21.266
77.919
87.78
826.185
20390.77
Inference Time (s) ↓
0.044
0.007
0.362
0.042
0.615
0.062
0.088
0.448
Table 3: Computational profile under the same output resolution. Trainable parameters exclude frozen pretrained priors where applicable.
Variant
Degradation-Control Signals
Quantitative Metrics
ℓd
hv
δ
Attr.
PSNR ↑
SSIM ↑
LPIPS ↓
ExpErr ↓
ColorErr ↓
Base model
✗
✗
✗
✗
21.358
0.7529
0.2796
0.0531
0.1614
Full DeLang-SR @5K
✓
✓
✓
✓
22.986
0.7594
0.2482
0.0360
0.1120
w/o degradation language
✗
✓
✓
✓
22.433
0.7583
0.2786
0.0383
0.1236
w/o visual degradation tokens
✓
✗
✓
✗
22.398
0.7500
0.2589
0.0423
0.1270
w/o descriptor conditioning in Mϕ
✓
✓
✗
✓
22.754
0.7577
0.2534
0.0335
0.1023
Table 4: Controlled ablation on RELLISUR ×2 under the same 5K-step training budget. ℓd , hv , and δ denote degradation language, visual degradation tokens, and descriptor conditioning of the update-strength predictor, respectively.
Figure 7: Qualitative ablation among different cases in Table 4 . ROI crops and error maps show that visual tokens and attribute alignment improve local texture recovery.
Figure 8: Checkpoint sensitivity on RELLISUR ×2 . Metrics become stable after 15K iterations, indicating that the final performance is not tied to a specific checkpoint.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Type
Statistic
Definition
Interpretation
P
Δe
Mean(Y(x))−Mean(Y(y))
Exposure gap between the target and low-light input.
P
Δc
Std(Y(x))−Std(Y(y))
Contrast degradation caused by insufficient illumination.
Table 5: Low-light degradation statistics used for degradation-condition construction. P and I denote paired-reference and input-only statistics, respectively.
Degradation Factor
Continuous Descriptor
Discrete Label / Slot
Role
Underexposure
δexp←Δˉe
κexp
Exposure intent and visibility recovery.
Dark regions
δdark←rdark
–
Local darkness-aware magnitude control.
Shadow noise
δnoise←nˉshadow
κnoise
Noise-aware restoration intent.
Contrast loss
δcontrast←Δˉc
κcontrast
Contrast restoration intent.
Detail degradation
δdetail←Δˉd
κdetail
Detail recovery intent.
Illumination non-uniformity
δillum←1nonuniform
κillum
Non-uniform illumination control.
Appendix
Table 6: Continuous descriptor entries and discrete degradation labels. The descriptor δ guides update-strength prediction, while κ fills degradation-language prompt slots and supervises attribute alignment.
Group
Term
Definition
Role
Latent alignment
Llat
∥zy−zM∥22
Aligns the degraded latent with the spatially modulated flow state.
Pixel reconstruction
Lpix
∥x^−x∥1
Supervises pixel-level fidelity.
Frequency reconstruction
Lfreq
∥Low(x^)−Low(x)∥1+∥High(x^)−High(x)∥1
Constrains illumination/color and high-frequency details.
Perceptual consistency
Lperc
DISTSDINOv3(x^,x)
Uses DISTS-style supervision with DINOv3 features [ 6 , 40 ] .
Update-strength prior
LM
∥Mϕ(y,δ)−Mprior(y)∥1
Calibrates the spatial update-strength field.
Attribute alignment
Lattr
∑k=1KCE(κ^k,κk)
Aligns visual tokens with discrete degradation labels.
Appendix
Table 7: Summary of training objectives used in DeLang-SR. The losses supervise latent flow alignment, image reconstruction, and degradation-control regularization.
(δ,κ,ℓd)=V(y,x).
Appendix
Algorithm 1 Training procedure of DeLang-SR
(δ,ℓd)=Vtest(y).
Appendix
Algorithm 2 Inference procedure of DeLang-SR
Method
LOL-v2-Real
LSRW
PSNR ↑
SSIM ↑
LPIPS ↓
DISTS ↓
ExpErr ↓
ColorErr ↓
PSNR ↑
SSIM ↑
LPIPS ↓
DISTS ↓
ExpErr ↓
ColorErr ↓
ASID
17.879
0.6683
0.4921
0.2721
0.0746
0.2264
17.293
0.5173
0.6048
0.2607
0.0613
0.1936
CATANet
17.488
0.6536
0.4567
0.2710
0.0861
0.2641
16.922
0.5127
0.5906
0.2650
0.0816
0.2580
HiT-SR
18.943
0.6985
0.4213
0.2863
0.0854
0.2667
16.174
0.5147
0.6061
0.2879
0.0908
0.2940
PFT-SR
18.927
0.6826
0.5298
0.2966
0.0844
0.2577
17.463
0.5098
0.6545
0.2833
0.0767
0.2378
DHGM
19.103
0.6189
0.4914
0.2789
0.0597
0.1866
17.383
0.5192
0.5036
0.2611
0.0561
0.1805
Appendix
Table 8: Full-reference cross-dataset evaluation under the constructed ×2 dark-SR protocol. All models are trained on RELLISUR and evaluated without fine-tuning.
Method
LOL-v2-Real
LSRW
MUSIQ ↑
MANIQA ↑
CLIP-IQA ↑
MUSIQ ↑
MANIQA ↑
CLIP-IQA ↑
ASID
38.078
0.2155
0.1925
38.064
0.2155
0.1684
CATANet
41.756
0.2088
0.1801
39.027
0.2008
0.1489
HiT-SR
32.061
0.2014
0.1458
33.438
0.2336
0.1508
PFT-SR
28.927
0.1770
0.1896
31.787
0.2285
0.1741
DHGM
43.575
0.2554
0.1919
43.127
0.2247
0.1667
Appendix
Table 9: No-reference perceptual evaluation under the same cross-dataset protocol.
Bounds
PSNR ↑
SSIM ↑
LPIPS ↓
ExpErr ↓
ColorErr ↓
[0.10,0.90]
22.910
0.7588
0.2536
0.0374
0.1148
[0.15,0.85]
22.880
0.7596
0.2551
0.0341
0.1065
[0.15,0.95] (default)
22.986
0.7594
0.2482
0.0360
0.1120
[0.20,0.90]
22.930
0.7601
0.2518
0.0349
0.1082
Appendix
Table 10: Inference-time sensitivity to (σmin,σmax) on RELLISUR ×2 at 5K steps.
Update strength
PSNR ↑
SSIM ↑
LPIPS ↓
ExpErr ↓
ColorErr ↓
Uniform image-wise mean
22.743
0.7541
0.2607
0.0358
0.1105
Spatial field
22.986
0.7594
0.2482
0.0360
0.1120
Appendix
Table 11: Ablation of spatial variation in the update-strength field on RELLISUR ×2 at 5K steps.
Figure 9: Additional visualizations of the update-strength field Mϕ(y,δ) . For each example, ℓd and δ are estimated from the input y without GT annotations. The predicted maps provide spatially non-uniform step-size fields for one-step SD3 flow restoration.
Figure 10: Additional ring-chart visualization of selected component ablations on RELLISUR ×2 . The chart compares Full DeLang-SR, w/o degradation language, and w/o descriptor conditioning in Mϕ across PSNR, SSIM, LPIPS, ExpErr, and ColorErr. The full model achieves a better overall trade-off, supporting the complementarity of direction steering and magnitude control.
Condition Source
Uses GT?
PSNR ↑
SSIM ↑
LPIPS ↓
ExpErr ↓
ColorErr ↓
Generic prompt
No
22.984
0.7593
0.2487
0.0361
0.1120
LQ-only degradation intent
No
22.985
0.7594
0.2482
0.0361
0.1120
Oracle degradation intent
Yes
22.985
0.7593
0.2479
0.0360
0.1119
Appendix
Table 12: Inference-time condition source analysis on RELLISUR ×2 . Comparable results across generic, LQ-only, and oracle intents show that DeLang-SR does not rely on oracle degradation annotations.
Figure 11: Additional visual comparison on RELLISUR ×2 . Each row shows the restored full image with the selected ROI marked in green and the corresponding enlarged crop. DeLang-SR preserves clearer local structures and more faithful contrast under challenging dark ×2 super-resolution.
Figure 12: Additional visual comparison on RELLISUR ×4 . Each row shows the full restored image and the enlarged ROI crop. DeLang-SR preserves clearer local structures and more faithful contrast under challenging dark ×4 super-resolution.
Despite recent advances, single-image super-resolution (SR) remains challenging, especially in real-world scenarios with complex degradations. Diffusion-based SR methods, particularly those built on Stable Diffusion, leverage strong generative priors but commonly rely on text conditioning derived from semantic captioning. Such textual descriptions provide only high-level semantics and lack the spatially aligned visual information required for faithful restoration, leading to a representation gap between abstract semantics and spatially aligned visual details. To address this limitation, we propose GramSR, a one-step diffusion-based SR framework that replaces text conditioning with dense visual features extracted from the low-resolution input using a pre-trained DINOv3 encoder. GramSR adopts a three-stage LoRA architecture, where pixel-level, semantic-level, and texture-level LoRA modules are trained sequentially. The pixel-level module focuses on degradation removal using ℓ2 loss, the semantic-level module enhances perceptual details via LPIPS and CSD losses, and the texture-level module enforces feature correlation consistency through a Gram matrix loss computed from DINOv3 features. At inference, independent guidance scales enable flexible control over degradation removal, semantic enhancement, and texture preservation. Extensive experiments on standard SR benchmarks demonstrate that GramSR consistently outperforms existing one-step diffusion-based methods, achieving superior structural fidelity and texture realism. The code for this work is available at: https://github.com/aimagelab/GramSR.
Generative image super-resolution (ISR) methods typically sample from a standard Gaussian prior and transport the sample to the data distribution, where the low-quality (LQ) input serves only as an external condition. However, the LQ image already carries rich structural information about the target high-quality (HQ) image, which makes a distribution approximating the LQ image a more suitable starting point, an idea that has been validated by a series of diffusion-based restoration methods. In contrast, recent flow-matching-based ISR methods still adopt the standard Gaussian prior and thereby discard such structural priors; moreover, existing single-step acceleration techniques, such as distillation and consistency training, either depend on an additional teacher or sacrifice the model's multi-step inference capability. In this paper, we propose Residual Flow Matching for Image Super-Resolution (RFMSR), which defines the source distribution as a Gaussian centered at the LQ latent, so that the flow starts from the LQ image rather than from pure noise and preserves the structural priors already present in the input. To achieve high-quality single-step generation, we further propose a simple yet effective training strategy that combines the supervision loss on the single-step output with the conditional flow matching (CFM) loss, enabling high-quality restoration with an arbitrary number of steps without resorting to distillation or other acceleration techniques. Extensive experiments show that RFMSR achieves competitive results against state-of-the-art methods. The source code is available at https://github.com/Faze-Hsw/RFMSR.
Single image super-resolution aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs. Training SR models typically requires paired HR-LR data, which is difficult to obtain in reality. As a result, most methods synthesize LR images by artificially degrading HR images with handcrafted kernels or camera ISP adjustments. However, these synthetic degradations fail to capture the complexity of real LR images, leading to poor generalization in practice. To address this, we observe that even within a single high-quality image, regions at different depths exhibit varying resolutions, where distant regions act as LR patches and closer ones as HR patches. This allows the extraction of real, degradation-induced LR patches from real images. Since these LR patches lack paired HR counterparts, we propose LA-SR (Language Assistant for SR), a novel framework for unpaired SR. The key idea of LA-SR is to redefine unpaired SR in the language space, using vision-language models to bridge the LR-HR gap. LA-SR projects images into a semantically rich space representing both content and quality, and applies two language-guided losses: linguistic content loss to preserve semantic fidelity, and linguistic quality loss to enhance perceptual realism. With this alignment, LA-SR effectively super-resolves real LR inputs, producing realistic outputs that overcome the limitations of synthetic-data-trained methods.
Joonkyu Park, Kyoung Mu Lee
Dept. of ECE&ASRI, 2IPAI, Seoul National University