Scene Text Image Super-Resolution (STISR) aims to restore high-resolution details in low-resolution text images, which is crucial for both human readability and machine recognition. Existing methods, however, often depend on external Optical Character Recognition (OCR) models for textual priors or rely on complex multi-component architectures that are difficult to train and reproduce. In this paper, we introduce DualTSR, a unified end-to-end framework that addresses both issues. DualTSR employs a single multimodal transformer backbone trained with a dual diffusion objective. It simultaneously models the continuous distribution of high-resolution images via Conditional Flow Matching and the discrete distribution of textual content via discrete diffusion. This shared design enables visual and textual information to interact at every layer, allowing the model to infer text priors internally instead of relying on an external OCR module. Compared with prior multi-branch diffusion systems, DualTSR offers a simpler end-to-end formulation with fewer hand-crafted components. Experiments on synthetic Chinese benchmarks and a curated real-world evaluation protocol show that DualTSR achieves strong perceptual quality and text fidelity.
Figures & tables
Figure 1: A comparison of text super-resolution architectures. Our proposed method unifies image generation and text recognition within a single model. (a) Baseline: Directly models the conditional distribution of high-resolution images from low-resolution inputs, without a text-based prior. (b) DiffTSR [ 44 ] : Independently models the image and text priors, subsequently fusing their representations using a Multi-modality Module (MoM). (c) Proposed Method: A unified multimodal transformer jointly optimizes for image generation and text recognition through two distinct but interconnected diffusion processes.
Figure 2: Illustration of the joint attention mechanism. Following the MM-DiT block design in SD3 [ 6 ] , we incorporate a joint attention module that enables the image latent refinement and text reconstruction processes to operate in parallel.
Method
× 2
× 4
PSNR ↑
LPIPS ↓
FID ↓
ACC ↑
NED ↑
PSNR ↑
LPIPS ↓
FID ↓
ACC ↑
NED ↑
ESRGAN [ 38 ]
24.09
0.3046
15.06
67.08%
83.79%
22.18
0.3986
18.25
43.69%
62.15%
MSRResNet [ 42 ]
28.03
0.3030
30.47
68.94%
85.37%
24.52
0.4029
50.60
49.04%
67.96%
SwinIR [ 16 ]
28.39
0.2983
31.54
70.01%
86.10%
24.73
0.3957
50.89
50.09%
68.93%
SRFormer [ 47 ]
28.89
0.2829
26.93
70.93%
86.91%
25.05
0.3801
46.23
51.83%
70.70%
MARCONet [ 13 ]
23.14
0.4518
88.06
57.93%
76.43%
25.05
0.3801
46.23
51.83%
70.70%
Table 1: Quantitative comparison on the synthetic CTR-TSR . Best in bold .
Method
×2
×4
PSNR ↑
LPIPS ↓
FID ↓
ACC ↑
NED ↑
PSNR ↑
LPIPS ↓
FID ↓
ACC ↑
NED ↑
ESRGAN [ 38 ]
20.08
0.3833
103.50
58.80%
86.30%
19.97
0.3721
87.18
58.50%
87.65%
MSRResNet [ 42 ]
19.25
0.3033
43.82
62.80%
89.08%
19.28
0.3590
60.81
58.00%
86.87%
SwinIR [ 16 ]
20.24
0.3810
87.98
59.30%
86.44%
20.23
0.3271
56.77
62.10%
87.55%
SRFormer [ 47 ]
20.85
0.3081
50.81
62.90%
89.47%
20.29
0.3732
83.47
57.50%
85.79%
MARCONet [ 13 ]
18.93
0.4144
87.54
60.58%
87.11%
20.30
0.3732
83.47
57.50%
85.79%
Table 2: Quantitative comparison on RealCE . Best in bold .
Figure 3: Qualitative comparison for the synthetic dataset CTR-TSR-Test with different methods on × 4 scale. The comparison methods include ESRGAN [ 38 ] , MSRResNet [ 42 ] , SwinIR [ 16 ] , SRFormer [ 47 ] , MARCONet [ 13 ] , MARCONet++ [ 14 ] , DiffTSR [ 44 ] and our method.
Figure 4: Qualitative comparison for RealCE with different methods on × 4 scale. The comparison methods include ESRGAN [ 38 ] , MSRResNet [ 42 ] , SwinIR [ 16 ] , SRFormer [ 47 ] , MARCONet [ 13 ] , MARCONet++ [ 14 ] , DiffTSR [ 44 ] and our method.
Setting
FID ↓
PSNR ↑
LPIPS ↓
ACC ↑
NED ↑
(a) LJoint-MG
13.73
19.82
0.3804
49.85%
68.28%
(b) LJoint-MG+LTXT
13.39
19.79
0.3782
52.72%
72.17%
(c) LJoint-MG+LTXT+LIMG-MG
9.92
20.12
0.3550
53.71%
73.60%
Table 3: Loss ablation study. All models are trained for 300k iterations with a batch size of 128.
Figure 5: Trade-off between perceptual quality and text fidelity across sampling steps. FID improves with more sampling steps, reaching its best value around ∼40 , while NED worsens as text strokes become over-smoothed. We select 4 steps as a balanced operating point.
Figure 6: Trade-off between perceptual quality and text fidelity across sampling steps.
CFG
1
1.45
2
4
6
FID ↓
16.10
19.59
24.19
35.35
39.49
NED ↑
76.66%
76.71%
75.95%
73.09%
72.20%
Table 4: Effect of CFG scale on FID and NED. Increasing CFG worsens FID and eventually harms NED.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Feature
DiffTSR
Ours
Training Stages
3-stage training
End-to-end training
Core Components
3 (IDM, TDM, MoM)
1 (Multimodal Transformer)
Auxiliary Models
Pre-trained VAE + text recognizer
Pre-trained VAE
Appendix
Table 5: Comparison of training framework. Our framework consolidates text recognition and super-resolution into a single, end-to-end module, whereas DiffTSR [ 44 ] requires a multi-stage pipeline with multiple components and auxiliary models.
Figure 7: Visual comparison for the synthetic dataset CTR-TSR-Test with different methods on ×2 scale. The comparison methods include ESRGAN [ 38 ] , MSRResNet [ 42 ] , SwinIR [ 16 ] , SRFormer [ 47 ] , MARCONet [ 13 ] , MARCONet++ [ 14 ] , DiffTSR [ 44 ] and our method.
Figure 8: Visual comparison for the synthetic dataset CTR-TSR-Test with different methods on ×2 scale. The comparison methods include ESRGAN [ 38 ] , MSRResNet [ 42 ] , SwinIR [ 16 ] , SRFormer [ 47 ] , MARCONet [ 13 ] , MARCONet++ [ 14 ] , DiffTSR [ 44 ] and our method.
Figure 9: Visual comparison for the synthetic dataset CTR-TSR-Test with different methods on ×4 scale. The comparison methods include ESRGAN [ 38 ] , MSRResNet [ 42 ] , SwinIR [ 16 ] , SRFormer [ 47 ] , MARCONet [ 13 ] , MARCONet++ [ 14 ] , DiffTSR [ 44 ] and our method.
Figure 10: Visual comparison for the real-world dataset RealCE with different methods on ×2 scale. The comparison methods include ESRGAN [ 38 ] , MSRResNet [ 42 ] , SwinIR [ 16 ] , SRFormer [ 47 ] , MARCONet [ 13 ] , MARCONet++ [ 14 ] , DiffTSR [ 44 ] and our method.
Figure 11: Visual comparison for the real-world dataset RealCE with different methods on ×2 scale. The comparison methods include ESRGAN [ 38 ] , MSRResNet [ 42 ] , SwinIR [ 16 ] , SRFormer [ 47 ] , MARCONet [ 13 ] , MARCONet++ [ 14 ] , DiffTSR [ 44 ] and our method.
Figure 12: Visual comparison for the real-world dataset RealCE with different methods on ×4 scale. The comparison methods include ESRGAN [ 38 ] , MSRResNet [ 42 ] , SwinIR [ 16 ] , SRFormer [ 47 ] , MARCONet [ 13 ] , MARCONet++ [ 14 ] , DiffTSR [ 44 ] and our method.
Figure 13: Visual comparison for the real-world dataset RealCE with different methods on ×4 scale. The comparison methods include ESRGAN [ 38 ] , MSRResNet [ 42 ] , SwinIR [ 16 ] , SRFormer [ 47 ] , MARCONet [ 13 ] , MARCONet++ [ 14 ] , DiffTSR [ 44 ] and our method.
Scene text image super-resolution (STISR) aims to recover visually plausible appearance while preserving character semantics from degraded inputs. Existing STISR systems often rely on externally generated priors or separate image and text models, resulting in error propagation and costly multi-stage inference. We present DualTSR, a unified framework that formulates STISR as coupled continuous-discrete generation. Conditional flow matching restores continuous image latents, while absorbing-state discrete diffusion reconstructs text tokens. Both processes share a multimodal transformer backbone, allowing the evolving image and text states to interact throughout generation without an external OCR prior at inference. On CTR-TSR, DualTSR achieves the best FID, LPIPS, ACC, and NED among the compared methods at both X2 and X4. On an aligned RealCE subset, it obtains the best FID, ACC, and NED with competitive LPIPS. Compared with DiffTSR at X4, DualTSR improves ACC by 12.78 percentage points while reducing the parameter count from 1.23B to 203M and end-to-end latency from 13.3s to 132ms. These results establish DualTSR as an accurate and efficient method for STISR.
Axi Niu, Knag Zhang, Qingsen Yan +3
Northwestern Polytechnical University, Xi’an, China · Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Republic of Korea
Text image super-resolution (TSR) aims to recover visually faithful and readable text under unknown degradations. Existing diffusion-based methods typically rely on multi-step prediction of either the high-resolution image or its text prior, resulting in prohibitive computational cost and inference latency. More critically, an erroneous text prior may be repeatedly injected into the denoising process, causing image and text predictions to reinforce each other and progressively amplify an early recognition error into a sharp yet semantically incorrect character. To address these limitations, we propose TOLA, a Text-aware One-step Latent Adaptation framework without iterative image-text diffusion. TOLA consists of two key modules. First, a confidence-weighted text conditioning module constructs the semantic condition only once and suppresses unreliable OCR predictions before they contaminate image reconstruction. Second, a lightweight latent residual correction module explicitly estimates and corrects the structured residual errors to recover missing or distorted stroke details. Extensive experiments demonstrate our state-of-the-art performance across all evaluation metrics on both CTR-TSR-Test (×4) and RealCE-200 benchmarks. It is worth noting that our TOLA consistently surpasses existing diffusion-based TSR methods by at least 2.72 dB in PSNR on CTR-TSR-Test.
Yike Xu, Yue Shi, Yong Guo +1
Shanghai Jiao Tong University · University of Chinese Academy of Sciences · Shanghai AI Laboratory
Diffusion models have achieved remarkable success in image super-resolution (SR) by leveraging strong generative priors to enhance perceptual quality. However, for textual content, these priors often produce text hallucinations: non-existent or distorted text, which alter scene semantics and compromise model reliability. While recent methods focus on enhancing text generation, they do so indiscriminately, leaving hallucinations fundamentally unresolved. We argue the core problem is a lack of selective control. Therefore, we reformulate text SR as a selective generation problem: text is resolved only when reliable generation is expected, while other scene content is fully super-resolved. We introduce TGSR (Text-Grounded Super-Resolution), a test-time framework that progressively identifies and suppresses likely hallucinated text regions during generation. To enable this, we train dedicated detectors on a large-scale dataset curated from diverse diffusion-based SR outputs. TGSR consistently improves text fidelity across multiple backbones, increasing recognition precision by up to 17.68 points without degrading perceptual quality, underscoring the value of selective control in generative SR.
Zhiming Hu, Ran Zhang, Angela Ye +7
AIC-Toronto, Samsung Electronics · AMD · University of Toronto +2