Text-aware image super-resolution (TAISR) aims to recover high-resolution images from low-resolution inputs while preserving textual content. Existing methods treat this as a visual restoration problem, yet severely degraded text often cannot be read from local visual evidence alone. Humans instead read it by reasoning over image context, logical patterns, and prior knowledge. This raises a fundamental question: Does TAISR need reasoning? To answer this, we introduce two diagnostic tools. The first, ReasonText, is a human-curated benchmark that separates reasoning-required text from locally recoverable text. The second, Ground-Truth Text-Crop Accuracy (GTTCA), is a recognition-based metric that measures text restoration quality without text detection errors. Evaluating twelve recent super-resolution (SR) models, we find a consistent and substantial performance drop on reasoning-required text, revealing a systematic blind spot. Two observations point to a remedy. One is that text-conditioned SR models restore text substantially better when given ground-truth text. The other is that multimodal large language models (MLLMs) can read degraded text in low-resolution images through reasoning. We therefore propose Reasoning Transfer via Captioning (RTC), a training-free plug-in that uses an external MLLM to infer the degraded text and feeds the resulting caption to a text-conditioned SR model. RTC consistently improves all six text-conditioned SR models we evaluate, especially on reasoning-required text. Surprisingly, a real-world SR model with RTC can even outperform the TAISR model with RTC, suggesting that reasoning should be a core design axis for future TAISR.
Figures & tables
Figure 2: ReasonText addresses gaps in existing TAISR benchmarks. (a) SA-Text-test assigns difficulty at the image level via three degradation strengths, which does not indicate whether each instance requires reasoning. (b) ReasonText instead labels each instance with a perception-grounded difficulty level, separating reasoning-required from locally recoverable text. (c) ReasonText is constructed from human annotations, which remain reliable across diverse scenes. (d) ReasonText crops the image to preserve context needed to read reasoning-required text.
Group
Model
Fidelity
Perceptual
Text Restoration
Reference
No Reference
GTTCA
PSNR-Y ↑
SSIM-Y ↑
LPIPS ↓
DISTS ↓
MANIQA ↑
MUSIQ ↑
CLIP-IQA ↑
Overall ↑
Level 1 ↑
Level 2 ↑
HR (oracle)
80.0000
1.0000
0.0000
0.0000
0.4261
66.2367
0.5403
94.6813
96.2942
90.6911
A
Real-ESRGAN
24.0453
0.7694
0.2134
0.1719
0.4463
68.5397
0.4611
40.3167
52.9647
9.0268
SwinIR
24.7762
0.7859
0.1989
0.1626
0.4288
66.9483
0.4357
45.6760
59.4071
11.7066
B
DiffBIR
24.1097
0.7039
0.2591
0.1880
0.6127
71.8427
0.6417
54.9736
69.6123
18.7588
Table 1: Quantitative results of twelve real-world SR and TAISR models on ReasonText. Every model struggles to recover Level 2 (reasoning-required) text. For each metric, the top-three models are in black ( best , second-best , third-best plain) and the rest in gray . The bottom rows report Pearson/Spearman correlations between each metric and overall GTTCA across the twelve SR models. ↑ ( ↓ ): higher (lower) is better.
Group
Model
Fidelity
Perceptual
Text Restoration
Reference
No Reference
GTTCA
PSNR-Y ↑
SSIM-Y ↑
LPIPS ↓
DISTS ↓
MANIQA ↑
MUSIQ ↑
CLIP-IQA ↑
Overall ↑
Level 1 ↑
Level 2 ↑
HR (oracle)
80.0000
1.0000
0.0000
0.0000
0.4261
66.2367
0.5403
94.6813
96.2942
90.6911
B
SUPIR
24.2971
0.7098
0.2622
0.1947
0.4608
63.5998
0.5922
55.2984
68.5291
22.5670
+ GT
24.5912
0.7038
0.2732
0.1964
0.4349
60.9064
0.5714
69.7686
81.6534
40.3667
FaithDiff
24.2038
0.7363
0.2061
0.1623
0.5095
71.6540
0.5821
52.0097
65.1653
19.4640
Table 2: Injecting ground-truth text improves GTTCA for all six text-conditioned SR models. Each model is evaluated with its original caption and with the GT transcripts of its Level 1 and Level 2 instances inserted as a natural-language caption. In the + GT rows, blue / red indicate an improvement/decline over the original caption. ↑ ( ↓ ): higher (lower) is better.
Model
Overall ↑
Level 1 ↑
Level 2 ↑
PaddleOCR-VL-1.5
48.03
54.90
31.03
TESTR
56.44
61.29
44.43
GLM-OCR
59.32
66.08
42.60
Table 3: The reasoning prompt substantially improves MLLMs on reasoning-required (Level 2) text, where OCR models and the no-reasoning prompt lag far behind. (Left) Text recall of dedicated OCR models. (Right) Text recall of MLLMs under the no-reasoning / reasoning prompts. Best per column in bold . ↑ : higher is better.
Figure 3: Gemma 4 reads locally illegible text by reasoning over relevant cues. Top: given only the target region (the locally illegible "4"), Gemma 4 fails. Adding random unrelated patches does not help, but adding the related numeric column "1"–"8" restores the correct reading. Bottom: likewise, "WHEY" is read only when the banner "70% BETTER THAN WHEY" is visible. Blue and red backgrounds mark correct and incorrect readings.
Group
Model
Fidelity
Perceptual
Text Restoration
Reference
No Reference
GTTCA
PSNR-Y ↑
SSIM-Y ↑
LPIPS ↓
DISTS ↓
MANIQA ↑
MUSIQ ↑
CLIP-IQA ↑
Overall ↑
Level 1 ↑
Level 2 ↑
HR (oracle)
80.0000
1.0000
0.0000
0.0000
0.4261
66.2367
0.5403
94.6813
96.2942
90.6911
B
SUPIR
24.2971
0.7098
0.2622
0.1947
0.4608
63.5998
0.5922
55.2984
68.5291
22.5670
+ RTC
23.9534
0.7071
0.2480
0.1846
0.5202
66.8668
0.6412
67.3163
78.8483
38.7870
FaithDiff
24.2038
0.7363
0.2061
0.1623
0.5095
71.6540
0.5821
52.0097
65.1653
19.4640
Table 4: Quantitative comparison of text-conditioned SR models with and without RTC on ReasonText. RTC consistently improves text restoration for all six models. Baselines use each model’s default caption source. In the + RTC rows, blue / red indicate an improvement/decline over the baseline. ↑ ( ↓ ): higher (lower) is better.
Figure 4: Qualitative results of RTC across reasoning types. For Level 2 instances of each reasoning type, the top and bottom rows show the outputs of six text-conditioned SR models without and with RTC, respectively. With RTC, some models can turn incorrect text into correct text.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Table A : Reliability of the ReasonText difficulty levels. (a) Agreement, the fraction of annotators who attempted an instance and read it correctly, for the Level 1 and Level 2 instances assigned under our rule. The rows give the share of instances in each agreement range, followed by the mean and median. (b) Level counts re-derived by requiring at least one, at least half, more than half, or all of the attempting annotators to read an instance correctly. Agreement is high at both levels, and Level 2 does not shrink under stricter thresholds but grows from 709 to 885 under unanimity.
Group
Model
Baseline
+ RTC
Overall ↑
Level 1 ↑
Level 2 ↑
Overall ↑
Level 1 ↑
Level 2 ↑
A
Real-ESRGAN
41.0
53.9
10.7
–
–
–
SwinIR
46.4
60.1
14.2
–
–
–
B
DiffBIR
55.7
70.9
20.2
–
–
–
SUPIR
55.8
69.3
24.1
67.8
79.5
40.2
FaithDiff
52.5
65.9
20.8
61.6
74.8
30.7
Appendix
Table B : The conclusions hold under the majority rule. GTTCA on ReasonText with the levels re-derived under the majority ( ≥ 50%) rule of Table A (b), which gives 1,685 Level 1 and 717 Level 2 instances. All twelve models drop sharply on Level 2, RTC improves all six text-conditioned SR models, and DiT4SR + RTC outperforms UniT + RTC. Blue marks an improvement over the baseline.
Figure A : Computing GTTCA on a restored text instance. Each text instance is cropped around its GT polygon with padding, rotated and flipped with its GT orientation, and recognized by an OCR model. The recognized text is compared with the GT transcript under normalized exact match.
Group
Model
Overall
Level 1
Level 2
GTTCA
w/o RF
OCR
GTTCA
w/o RF
OCR
GTTCA
w/o RF
OCR
HR (oracle)
94.7
92.7
29.0
96.3
94.7
30.0
90.7
87.6
26.7
A
Real-ESRGAN
40.3
38.2
12.0
53.0
50.7
16.6
9.0
7.3
0.6
SwinIR
45.7
43.3
13.9
59.4
57.0
18.6
11.7
9.3
2.3
B
DiffBIR
55.0
53.5
15.1
69.6
67.8
20.0
18.8
17.9
3.0
SUPIR
55.3
54.0
16.8
68.5
67.1
21.6
22.6
21.4
4.9
Appendix
Table C : Both GT polygons and GT orientations are needed for GTTCA to reflect restoration alone. All three protocols share the same OCR backbone [ 6 ] . GTTCA uses both GT polygons and GT orientations, w/o RF uses GT polygons without rotation and flipping, and OCR detects and recognizes text end-to-end without either. The HR oracle row shows the ceiling of each protocol on images that need no restoration. Excluding the HR oracle, bold marks the best score in each column, and underline marks the second-best.
Group
Model
Baseline (Overall)
+ RTC (Overall)
Paddle
ASTER
MORAN
Paddle
ASTER
MORAN
A
Real-ESRGAN
40.3
25.8
26.8
–
–
–
SwinIR
45.7
30.9
30.8
–
–
–
B
DiffBIR
55.0
38.0
40.1
–
–
–
SUPIR
55.3
38.3
41.7
67.3
49.2
50.9
FaithDiff
52.0
36.4
38.3
61.0
43.6
45.0
Appendix
Table D : GTTCA under different OCR recognizers. Overall GTTCA recomputed with PaddleOCR-VL-1.5 ( Paddle ), ASTER, and MORAN. The weaker recognizers lower every score by a similar proportion, so the ranking of the models barely changes (Table E ). In the + RTC columns, blue marks an improvement over the baseline under the same recognizer.
Ranking (Overall GTTCA)
τ vs ASTER
τ vs MORAN
ρ vs ASTER
ρ vs MORAN
12 SR models
0.94
0.94
0.98
0.98
+ RTC (6 SR models)
1.00
1.00
1.00
1.00
Appendix
Table E : Ranking agreement across recognizers. Kendall τ and Spearman ρ between the model ranking under PaddleOCR-VL-1.5 and those under ASTER and MORAN, computed on Overall GTTCA in Table D . 1.0 denotes an identical ranking. Both measures stay near 1.0 , so GTTCA ranks the SR models in nearly the same order regardless of which recognizer reads the output.
Figure B : SR models fail on reasoning-required text. Outputs of the twelve SR models on one Level 1 (locally recoverable) instance and three Level 2 (reasoning-required) instances, one for each cue. Blue and red borders on the SR outputs mark correct and incorrect restorations. On Level 1, most models restore “Crab” correctly. On Level 2, most models fail to restore the text.
Group
Model
Fidelity
Perceptual
Text Restoration
Reference
No Reference
GTTCA
PSNR-Y ↑
SSIM-Y ↑
LPIPS ↓
DISTS ↓
MANIQA ↑
MUSIQ ↑
CLIP-IQA ↑
Overall ↑
Level 1 ↑
Level 2 ↑
HR (oracle)
80.0000
1.0000
0.0000
0.0000
0.4261
66.2367
0.5403
94.6813
96.2942
90.6911
B
SUPIR
24.2971
0.7098
0.2622
0.1947
0.4608
63.5998
0.5922
55.2984
68.5291
22.5670
+ tag style GT (space)
24.7121
0.7014
0.2840
0.2013
0.4332
60.6653
0.5659
64.9695
77.9590
32.8350
+ tag style GT (comma)
24.7246
0.7006
0.2852
0.2024
0.4291
60.2502
0.5620
64.6935
78.1984
31.2835
Appendix
Table F : The natural-language caption restores text best among the three caption formats. Each text-conditioned SR model is evaluated with its original caption and with the GT transcripts of its Level 1 and Level 2 instances injected in three formats. Within each model block, bold marks the best value for each metric, and blue / red indicate improvement/decline over the original caption. The HR oracle row reports each metric on the HR images for reference. ↑ ( ↓ ): higher (lower) is better.
Group
Model
Fidelity
Perceptual
Text Restoration
Reference
No Reference
GTTCA
PSNR-Y ↑
SSIM-Y ↑
LPIPS ↓
DISTS ↓
MANIQA ↑
MUSIQ ↑
CLIP-IQA ↑
Overall ↑
Level 1 ↑
Level 2 ↑
HR (oracle)
80.0000
1.0000
0.0000
0.0000
0.4261
66.2367
0.5403
94.6813
96.2942
90.6911
B
SUPIR
24.2971
0.7098
0.2622
0.1947
0.4608
63.5998
0.5922
55.2984
68.5291
22.5670
+ caption style GT
24.5912
0.7038
0.2732
0.1964
0.4349
60.9064
0.5714
69.7686
81.6534
40.3667
+ scene + caption style GT
23.9171
0.7059
0.2497
0.1828
0.5293
67.4045
0.6405
67.9009
79.0878
40.2257
Appendix
Table G : Adding a scene description to the caption-style GT largely recovers no-reference perceptual quality. Each text-conditioned SR model is evaluated with its original caption, with the caption-style GT of its Level 1 and Level 2 instances, and with a Gemma 4 scene description prepended to that caption. Within each model block, bold marks the best value for each metric, and blue / red indicate improvement/decline over the original caption. The HR oracle row reports each metric on the HR images for reference. ↑ ( ↓ ): higher (lower) is better.
Figure C : Overview of Reasoning Transfer via Captioning (RTC). (1) Caption by reasoning: a frozen MLLM reads the LR image under the reasoning prompt and returns a two-sentence caption with a scene description and the inferred transcripts. (2) Restore by conditioning: a frozen text-conditioned SR model restores the LR image with this caption in place of its default caption.
Group
Model
RealText
SA-Text-test
Level 1
Level 2
Level 3
HR (oracle)
75.4443
76.2919
76.2919
76.2919
B
SUPIR
41.6263
23.5721
20.2629
12.2847
+ RTC
50.9962
34.0435
28.4225
18.5857
FaithDiff
43.4572
23.5721
20.7162
13.3726
+ RTC
49.9731
28.4678
26.5186
18.7670
Appendix
Table H : RTC improves text restoration on existing TAISR benchmarks. Text-spotting metric (TESTR [ 46 ] ) measured on RealText and the three degradation levels of SA-Text-test, comparing each text-conditioned SR model with and without RTC. The levels of SA-Text-test denote degradation strengths and are unrelated to the difficulty levels of ReasonText. Blue marks an improvement over the baseline. The HR oracle row reports the score on the HR images for reference.
Table I : RTC is robust to caption errors. (a) Level 2 GTTCA before (base) and after RTC (+RTC), split by whether the RTC caption contains the correct GT transcript. RTC improves GTTCA for all six SR models even when Gemma 4 is wrong. (b) The Gemma wrong subset pooled over the six SR models and split by how far the caption is from the GT transcript. The gain concentrates on near-miss captions that differ by at most two letters, from which the SR models can still restore the exact text. Blue marks an improvement over the baseline.
Table J : RTC is effective at every difficulty level, including Level 3. (a) GTTCA before (base) and after RTC (+RTC) on each difficulty level. RTC improves all six SR models on Level 1, Level 2, and even Level 3, though the Level 3 gains are smaller on average. Mean Δ is the average gain over the six models. (b) Text recall of Gemma 4 with the reasoning prompt on each level. Blue marks an improvement over the baseline.