cs.CVOct 9, 2026
SaveTGSR: Text-Grounded Diffusion for Faithful Image Super-Resolution
Organizations: AIC-Toronto, Samsung Electronics · AMD · University of Toronto · Vector Institute · York University
Abstract
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.
Figures & tables
Figure 1: Comparison of state-of-the-art approaches with TGSR (ours) on text-aware image super-resolution (SR) . Existing SR models often produce distorted text when the corresponding regions in the low-resolution (LR) image are heavily degraded, as shown in the red bounding box. Our approach extends existing SR models ( e.g . , TeReDiff ( Min et al. 2026 ) ) to suppress hallucinated text artifacts at inference time without sacrificing the visual quality of legible text and background regions.
Figure 2: Dataset curation pipeline. From each ground-truth (GT) image, we generate three degradation levels using the Real-ESRGAN degradation adopted in prior works ( i.e . , TeReDiff, StableSR, PASD). The degraded images are super-resolved using four diffusion-based SR models: TeReDiff, PiSA-SR, PASD, and SUPIR. We then crop text regions based on the GT bounding boxes and recognize them using an MLLM-based OCR model. If the recognized text matches with GT transcription, the sample is labeled as normal text ; otherwise, it is labeled as hallucinated text .
| Dataset | Task | Text | Halluci. | # of images |
|---|---|---|---|---|
| Label | Label | |||
| SA-Text | Text-Aware SR | ✓ | ✗ | 105,330 |
| CTW1500 | Text Spotting | ✓ | ✗ | 1,000 |
| TotalText | Text Spotting | ✓ | ✗ | 1,255 |
| SA-Text-HTD | Halluci. Text Det. | ✓ | ✓ | 12 105,330 |
Table 1: Comparison with other datasets. Halluci. denotes hallucination . Text-aware SR and text spotting datasets ( Yuliang et al. 2017 ; Ch’Ng and Chan 2017 ) provide text labels but lack hallucination labels. Our dataset includes both text and fine-grained hallucination labels. Moreover, we enlarge the SA-Text ( Min et al. 2026 ) dataset by a factor of .
Figure 3: Overall architecture of TGSR . The framework contains two key components: a hallucinated text detector and test-time restoration guidance . The detector reuses the multi-scale features from the SR model’s own frozen denoising U-Net: only a lightweight text-spotting head is trained, which outputs text polygons with binary labels indicating whether each text instance is normal or hallucinated. During inference, at each diffusion step, we first predict the clean latent through denoising, and apply the detector head to obtain hallucinated text masks. The masked regions are used to compute the loss, and a one-step gradient update is applied to the clean latent before predicting . This process repeats for all diffusion steps.
| PSNR | SSIM | LPIPS | MUSIQ | CLIPIQA | QAlign | OCR-P | OCR-R | OCR-F1 | |
| Bicubic | |||||||||
| Swin2SR ( Conde et al. 2022 ) | |||||||||
| Real-ESRGAN ( Wang et al. 2021 ) | |||||||||
| StableSR ( Wang et al. 2024 ) | |||||||||
| PASD ( Yang et al. 2024 ) | |||||||||
| SeeSR ( Wu et al. 2024c ) |
Table 2: SR results on Real-Text dataset ( Min et al. 2026 ) . We evaluate on FR-IQA, NR-IQA, and end-to-end text recognition metrics. We compare base SR models and our adapted models with test-time adaptation, as denoted by +TGSR (ours) . Our method increases recognition precision (OCR-P) and OCR-F1 while maintaining the perceptual quality.
Figure 4: Qualitative results on the Real-Text dataset . Our method suppresses text hallucinations (top) and preserves background details (bottom) compared to other diffusion-based SR models.
| TeReDiff | SeeSR † | StableSR † | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Detector | P | R | F1 | P | R | F1 | P | R | F1 |
| TESTR | 28.2 | 47.7 | 35.4 | 24.2 | 34.4 | 28.4 | 23.5 | 30.4 | 26.5 |
| TeReDiff | 25.2 | 55.9 | 34.7 | 25.2 | 47.4 | 33.0 | 26.0 | 44.9 | 32.9 |
| Ours (RN50) | 53.0 | 32.6 | 40.4 | 57.2 | 28.7 | 38.2 | 60.4 | 26.9 | 37.2 |
| Ours (DF) | 60.1 | 32.6 | 42.3 | 62.4 | 24.7 | 35.4 | 65.3 | 25.5 | 36.7 |
Table 3: Hallucinated text detection on Real-Text-HTD. Precision (P) and recall (R) on hallucinated regions at the deployed operating point ( ); marks SR backbones unseen during detector training. In the context of test-time adaptation, false positives may erroneously blur sharp text by pulling it toward the reference image, whereas false negatives simply result in a no-op for that region. Consequently, we prioritize precision in the hallucinated text detector to avoid degrading existing high-quality textual content.
| OCR-P | OCR-R | OCR-F1 | ||||
|---|---|---|---|---|---|---|
| Text size | Base | +TGSR | Base | +TGSR | Base | +TGSR |
| Small (p25–p50) | 40.2 | 64.9 | 46.0 | 44.9 | 42.9 | 53.1 |
| Ext. small ( p25) | 15.0 | 36.5 | 12.3 | 10.8 | 13.5 | 16.6 |
Table 4: Recognition by text size on Real-Text (via GT polygon area percentile), for TeReDiff (Base) and with TGSR. Gains are largest exactly where hallucination is most frequent: on extremely small text, precision more than doubles.
| NR-IQA | OCR | ||||
| Configuration | MUSIQ | CLIPIQA | P | R | F1 |
| TeReDiff | 62.37 | 0.643 | 53.24 | 42.43 | 47.23 |
| Ours | 67.24 | 0.697 | 70.92 | 42.16 | 52.89 |
| Guidance space and objective | |||||
| RGB, MSE | 60.93 | 0.627 | 68.60 | 42.11 | 52.19 |
| Latent, MSE | 61.39 | 0.631 | 70.93 | 41.79 | 52.59 |
Table 5: Ablation studies on Real-Text using TeReDiff as the base model. Our default configuration employs a masked latent MSE objective, incorporates MUSIQ ( ) and CLIPIQA ( ) on decoded images, and applies a dampening factor of for empty masks.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
| SR Models | ||||
|---|---|---|---|---|
| Real-ESRGAN Degradations | TeReDiff | PiSA-SR | PASD | SUPIR |
| TeReDiff ( Min et al. 2026 ) | 215,489 | 225,938 | 223,471 | 229,019 |
| StableSR ( Wang et al. 2024 ) | 230,040 | 235,771 | 233,170 | 238,013 |
| PASD ( Yang et al. 2024 ) | 199,038 | 211,329 | 208,058 | 214,275 |
Table 6: Number of hallucinated text instances. Each HR image corresponds to 12 SR images (three degradation levels and four SR models). This table reports the number of hallucinated text regions for each combination of degradation level and SR model. Overall, among the 271,226 total text crops, the majority correspond to the hallucinated text regions. Across degradation levels, fewer hallucinated text regions are produced when images are degraded using the PASD ( Yang et al. 2024 ) degradation, likely because the SR models used in dataset curation are typically trained with degradation settings similar to those in PASD. Across models, TeReDiff ( Min et al. 2026 ) produces the fewest hallucinated text regions, highlighting the advantage of text-aware SR.
Figure 5: Example patches from SA-Text-HTD (ours) dataset. Most SR models recover text that is large and readable in the LR image despite the degradation ( e.g . , “TEAM”). Conversely, we observe a higher amount of hallucination in the SR output of small non-readable text ( e.g . , “Yonge”) and large non-readable text ( e.g . , “brandedcities”). We do not visualize the remaining readable text ( i.e . , “HALEON”, “E” and “ADVIL”) since their SR outputs are nearly identical and non-hallucinatory.
| Inference cost | Image quality | Recognition | ||||||
|---|---|---|---|---|---|---|---|---|
| Method | Time (s) | Mem (GB) | PSNR | LPIPS | MUSIQ | QAlign | OCR-F1 | |
| TeReDiff | – | 0 7.9 | 24.98 | 0.286 | 62.37 | 3.551 | 47.23 | |
| w/ RGB MSE only | 14.6 | 25.27 | 0.284 | 60.93 | 3.430 | 52.19 | ||
| +TGSR-lite | 0 7.8 | 25.17 | 0.283 | 61.39 | 3.446 | 52.59 | ||
| +TGSR | 15.7 | 25.05 | 0.286 | 67.24 | 3.534 | 52.89 | ||
| PiSA-SR | – | 0 9.6 | 26.46 | 0.241 | 64.37 | 3.260 | 46.86 | |
Table 7: Inference cost of different models. Time is the mean wall-clock time to restore one image and Mem is the peak GPU memory that restoration allocates, both measured on a single GPU with images processed one at a time; subscripts are the standard deviation across images. is the extra time over the base model of the same block, w/ RGB MSE only is an alternative to our default, and TGSR-lite is the variant that keeps only the masked latent MSE term of the guidance objective. PiSA-SR was profiled on one NVIDIA RTX 6000 Ada (48 GB) over all 847 Real-Text images, whereas TeReDiff and SUPIR were profiled on one NVIDIA H100 (80 GB) over 100 images.
Figure 6: Illustration of three levels of degradations in SA-Text-Test dataset . The task of producing faithful super-resolution for the text region becomes increasingly difficult with higher levels of degradation.
Figure 7: Qualitative results on the Real-Text dataset . We show two examples. For each example, the first image is the GT image with a bounding box defining a cropped region containing text, and the rest are the crops in GT and outputs from the various SR methods. When the input image lacks sufficient visual cues, diffusion-based SR models tend to hallucinate text in the restored regions. Our approach (TeReDiff + Ours) effectively suppresses these hallucinations while preserving plausible structures.
Figure 8: Qualitative results on the SA-Text-Test dataset . We show three examples. For each example, the first image is the GT image with a bounding box defining a cropped region containing text, and the rest are the crops in GT and outputs from the various SR methods. When the input image lacks sufficient visual cues, diffusion-based SR models tend to hallucinate text in the restored regions. Our approach (TeReDiff + Ours) effectively suppresses these hallucinations while preserving plausible structures.
Figure 9: Qualitative results on the Real-Text (top two rows) and SA-Text-Test (bottom three rows) datasets . For each example, the first image is the GT image with a bounding box defining a cropped region containing text, and the rest are the crops in GT and outputs from the TeReDiff and our two models. Compared to TeReDiff, our approaches (TeReDiff + Ours(Lite)) and (TeReDiff + Ours) effectively suppress hallucinations while preserving plausible structures. Our lite model (TeReDiff + Ours(Lite)) exhibits competitive performance.