Most of the recent generative image super-resolution (SR) methods rely on adapting large text-to-image (T2I) diffusion models pretrained on web-scale text-image data. While effective, this paradigm starts from a generic T2I generator, despite that SR is fundamentally a low-resolution (LR) input-conditioned image restoration task. In this work, we investigate whether an SR model trained purely on visual data can rival T2I-based ones. To this end, we propose VOSR, a Vision-Only generative framework for SR. We first extract semantically rich and spatially grounded features from the LR input using a pretrained vision encoder as visual semantic guidance. We then revisit classifier-free guidance for training generative models and show that the standard unconditional branch is ill-suited to restoration models trained from scratch. We therefore replace it with a restoration-oriented guidance strategy that preserves weak LR anchors. Built upon these designs, we first train a multi-step VOSR model from scratch and then distill it into a one-step model for efficient inference. VOSR requires less than one-tenth of the training cost of representative T2I-based SR methods, yet in both multi-step and one-step settings, it achieves competitive or even better perceptual quality and efficiency, while producing more faithful structures with fewer hallucinations on both synthetic and real-world benchmarks. Our results, for the first time, show that high-quality generative SR can be achieved without multimodal pretraining. The code and models can be found at https://github.com/cswry/VOSR.
Figures & tables
Figure 1: Comparison of VOSR with existing generative SR methods in terms of performance , efficiency , and training cost . Blue/orange colors denote T2I-based and vision-only methods, circles/triangles denote multi-step and one-step models. Performance is measured on RealSR [ 2 ] , and efficiency is measured at 512×512 resolution using official repositories. VOSR achieves competitive or better perceptual quality than many T2I-based SR methods in both multi-step and one-step settings, while clearly outperforming prior vision-only methods. Its multi-step variant is substantially more efficient than existing T2I-based methods, and its one-step variant remains comparable to recent one-step T2I systems. Measured by the total number of training pixels consumed, VOSR requires only about one-tenth of the training cost of representative T2I-based SR methods. For fairness, we count only the pretraining cost of the core diffusion modules; reused components such as the VAE and semantic encoders are excluded.
Figure 2: Overview of VOSR. (a) Framework overview. Given an LR image, VOSR builds two complementary conditions from the input: a spatially aligned structural condition in the VAE latent space and a high-level visual semantic condition extracted by a pretrained vision encoder. These conditions are injected into a diffusion transformer to predict the denoising velocity for HR reconstruction. (b) Condition and guidance design. Compared with prior vision-only SR that mainly relies on structural-only conditioning, VOSR introduces an additional visual semantic condition to reduce semantic ambiguity in restoration. Moreover, instead of using a fully unconditional branch as in standard classifier-free guidance, VOSR adopts a restoration-oriented guidance and removes semantic guidance while retaining weakened LR structural cues, making the guidance input-anchored and more suitable for restoration.
Dataset
Setting
Method
Type
PSNR ↑
SSIM ↑
LPIPS ↓
DISTS ↓
AFINE-FR ↓
NIQE ↓
MUSIQ ↑
MANIQA ↑
AFINE-NR ↓
TOPIQ-NR ↑
LSDIR
Multi-step
StableSR
T2I
19.1120
0.3751
0.5169
0.3068
0.5477
5.4703
56.2263
0.5345
-0.7529
0.5627
PASD
19.0817
0.4186
0.4898
0.2433
-0.1689
4.4446
65.6496
0.5780
-0.9574
0.5827
SeeSR
19.1408
0.4198
0.4180
0.2147
-0.7413
4.2406
72.3981
0.6616
-1.0766
0.7316
DiT4SR
17.7220
0.3833
0.4458
0.2246
-0.7300
4.2398
73.3560
0.6854
-1.1859
0.6951
ResShift
VO
19.8723
0.4300
0.4784
0.2699
1.0205
6.0833
59.1250
0.5243
-0.7036
0.4966
VOSR-0.5B-ms
19.5677
0.4312
0.3984
0.2126
-0.8310
4.0466
73.1033
0.6679
-1.0844
0.7231
Table 1: Quantitative results on LSDIR, ScreenSR, and RealSR. Methods are grouped into multi-step (ms) and one-step (os) settings. T2I and VO denote text-to-image-based and vision-only methods, respectively. ↑ ( ↓ ) indicates higher (lower) is better. T2I-based methods are marked in gray for visual distinction. The best and second-best results are highlighted in bold red and bold blue , respectively.
Multi-step ⋅ Vision-only
Multi-step ⋅ T2I-based
One-step ⋅ Vision-only
One-step ⋅ T2I-based
ResShift
VOSR-0.5B-ms
VOSR-1.4B-ms
StableSR
PASD
SeeSR
DiT4SR
SinSR
VOSR-0.5B-os
VOSR-1.4B-os
OSEDiff
PiSA-SR
Params (M)
174.0
619.4
1,742.9
1,540
1,680
2,510
21,380
174.0
620.7
1,745.7
1,760
1,290
Time (s)
0.534
1.226
2.114
10.036
2.751
3.943
12.550
0.116
0.094
0.095
0.104
0.091
Table 2: Complexity comparison on 512×512 inputs. Runtime is measured on an A100 GPU with batch size 1 in FP16.
Figure 3: Multi-step (top) and one-step (bottom) SR visual comparison on RealDeg [ 6 ] Cf/0020.png and ScreenSR 010.png .
Figure 4: Effect of guidance scale on VOSR-1.4B-ms. As the scale increases, LPIPS rises while MUSIQ drops, indicating a shift from more generative to more faithful restoration to the LR input.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Method
LPIPS ↓
MUSIQ ↑
Teacher (VOSR-0.5B-ms)
0.3069
68.93
Shortcut-based distillation
0.2913
68.21
RC-based distillation
0.2856
69.78
Appendix
Table 3: Comparison of shortcut-based and RC-based distillation on RealSR using VOSR-0.5B-ms as the teacher.
Figure 5: Thumbnail montage of the ScreenSR benchmark. The selected 130 GT images cover diverse scenarios, including indoor and outdoor scenes, humans, animals, plants, artworks, and multilingual text, with substantial variation in object and scene scales. This diversity ensures a comprehensive evaluation of generative SR methods in terms of semantic coverage, structural fidelity, and robustness across different content types.
Benchmark
NIQE ↓
MUSIQ ↑
MANIQA ↑
AFINE-NR ↓
TOPIQ-NR ↑
ScreenSR
3.7719
72.1500
0.7187
-1.2093
0.7363
RealSR [ 2 ]
6.1167
57.4564
0.6016
-0.9088
0.4140
DRealSR [ 42 ]
6.7909
50.5644
0.5588
-0.7731
0.3932
Appendix
Table 4: No-reference quality comparison of GT images in different real-world paired SR benchmarks. Better GT quality is indicated by lower NIQE and AFINE-NR, and higher MUSIQ, MANIQA, and TOPIQ-NR.
PT 256px
PT 512px
Distill to One-step
Learning rate
1.0×10−4
5.0×10−5
2.0×10−5
LR scheduler
Constant
Constant
Constant
Warm-up steps
0
0
0
Training steps
400K
400K
50K
Global batch size
1024
256
32
Weight decay
0.01
0.01
0.01
Appendix
Table 5: Training recipe for VOSR, including the progressive pretraining for the multi-step models and the distillation for the one-step model.
Method
LPIPS ↓
MUSIQ ↑
w/o SVE
0.3011
63.74
w/ SVE (CLIP)
0.2788
63.82
w/ SVE (SigLIPv2)
0.2817
64.81
w/ SVE (DINOv3)
0.2858
67.51
w/ SVE (DINOv2)
0.2872
68.23
Appendix
Table 6: Ablation on visual semantic encoders on RealSR.
Figure 6: User study results in the multi-step and one-step settings. VOSR-1.4B-ms and VOSR-1.4B-os receive the highest numbers of votes in their respective groups, showing strong human preference in perceptual quality and consistency with the LR input.
Guidance design
LPIPS ↓
MUSIQ ↑
Full condition only
0.3752
67.29
Standard CFG
0.4053
50.78
Ours (partial conditioning)
0.3772
69.26
Appendix
Table 7: Ablation studies on restoration-oriented partial conditioning on LSDIR.
Figure 7: Additional visual comparisons of multi-step (1st, 3rd and 5th) and one-step (2nd, 4th and 6th) SR results. Compared with representative vision-only and T2I-based methods, VOSR produces perceptually more realistic details while better preserving structures that are faithful to the LR input. Please zoom in for better view.