The authors of “Reading the Unreadable” report that a training-free plug-in called RTC lifts text restoration in text-conditioned super-resolution models: for DiT4SR, Level 2 GTTCA rises from 24.54% to 45.84%, a gain of 21.30 points (arXiv). The paper also introduces ReasonText, a human-curated benchmark that separates text that needs reasoning from text that can be recovered locally (arXiv). In our reading, the result suggests that part of the text failure in restoration may come from what the caption lacks, not only from the restoration network.
How RTC works
RTC uses an external multimodal language model to infer the degraded text and write it into a caption. That caption replaces the default caption of a text-conditioned super-resolution model, and because both models stay frozen, no training is needed. According to the paper, Gemma 4 with the reasoning prompt reached 70.38% text recall on Level 2, which is 25.95 points above the best OCR model, and it was chosen as the reasoning generator.
To keep the comparison controlled, the authors compute a single caption for each low-resolution image in advance, give that identical caption to all six models and leave it unchanged during denoising. They also report that RTC improves text restoration for all six models on the existing SA-Text-test and RealText benchmarks under their original evaluation protocol. Even on the subset where Gemma was wrong, RTC improves GTTCA for all six models, according to the paper.
One comparison stands out: DiT4SR with RTC outperforms UniT with RTC on both levels, although both receive the same caption and both build on the same backbone, and UniT adds text-aware modules. In our reading, this suggests the caption matters more than extra text modules once the caption carries the right text.
What ground-truth captions do
| Group | Model | Text Restoration: GTTCA: Overall ↑ | Text Restoration: GTTCA: Level 1 ↑ | Text Restoration: GTTCA: Level 2 ↑ | Source page |
|---|---|---|---|---|---|
| HR (oracle) | HR (oracle) | 94.6813 | 96.2942 | 90.6911 | https://arxiv.org/html/2611.15449v1 |
| B | SUPIR | 55.2984 | 68.5291 | 22.5670 | https://arxiv.org/html/2611.15449v1 |
| B | + tag style GT (space) | 64.9695 | 77.9590 | 32.8350 | https://arxiv.org/html/2611.15449v1 |
| B | + tag style GT (comma) | 64.6935 | 78.1984 | 31.2835 | https://arxiv.org/html/2611.15449v1 |
| B | + caption style GT | 69.7686 | 81.6534 | 40.3667 | https://arxiv.org/html/2611.15449v1 |
| B | FaithDiff | 52.0097 | 65.1653 | 19.4640 | https://arxiv.org/html/2611.15449v1 |
| B | + tag style GT (space) | 61.5510 | 74.6522 | 29.1396 | https://arxiv.org/html/2611.15449v1 |
| B | + tag style GT (comma) | 61.2667 | 74.7320 | 27.9549 | https://arxiv.org/html/2611.15449v1 |
| B | + caption style GT | 62.4523 | 75.3250 | 30.6065 | https://arxiv.org/html/2611.15449v1 |
| B | DiT4SR | 59.1555 | 73.1471 | 24.5416 | https://arxiv.org/html/2611.15449v1 |
| B | + tag style GT (space) | 69.9959 | 82.5314 | 38.9845 | https://arxiv.org/html/2611.15449v1 |
| B | + tag style GT (comma) | 70.1340 | 82.8050 | 38.7870 | https://arxiv.org/html/2611.15449v1 |
| B | + caption style GT | 74.1048 | 85.1767 | 46.7137 | https://arxiv.org/html/2611.15449v1 |
| E | UniT | 61.3886 | 76.4538 | 24.1185 | https://arxiv.org/html/2611.15449v1 |
| E | + tag style GT (space) | 64.5635 | 78.7001 | 29.5910 | https://arxiv.org/html/2611.15449v1 |
| E | + tag style GT (comma) | 64.4417 | 78.5063 | 29.6474 | https://arxiv.org/html/2611.15449v1 |
| E | + caption style GT | 70.6618 | 82.8848 | 40.4231 | https://arxiv.org/html/2611.15449v1 |
Table F isolates the caption effect by injecting ground-truth transcripts in three formats, with the HR oracle row as a reference. The paper says the transcripts improve GTTCA over the original caption for every model, whatever the format, and that the natural-language caption gives the best GTTCA for every model on both levels. The authors hypothesize that the text-to-image models underlying these super-resolution models were pretrained on natural-language captions.
The shown rows bear this out for SUPIR: its Level 2 GTTCA is 22.5670, and it reaches 40.3667 with the caption style GT, against 32.8350 and 31.2835 for the two tag styles. FaithDiff follows the same pattern on Level 2: 19.4640 originally, 30.6065 with caption style GT, and 29.1396 and 27.9549 with the tag styles. For DiT4SR, Level 2 GTTCA goes from 24.5416 to 46.7137 with caption style GT, while the tag styles give 38.9845 and 38.7870. UniT shows the same ordering on Level 2: 24.1185 originally, 40.4231 with caption style GT, and 29.5910 and 29.6474 with the tag styles. Level 1 gaps are smaller: SUPIR moves from 68.5291 to 81.6534 with the caption style, while the HR oracle sits at 96.2942. Even with the best format, SUPIR (40.3667) and FaithDiff (30.6065) stay far below the HR oracle’s Level 2 score of 90.6911, so the format choice narrows only part of the gap.
Limits the authors state
ReasonText is limited in size and scope: it holds 513 LR–HR pairs of natural images, and only alphanumeric text is annotated. RTC applies only to text-conditioned super-resolution models and cannot be used with models that have no text channel. Even the best model with RTC stays far below the HR oracle on Level 2, at 45.84% versus 90.69%. The same model reaches at most 47.70% with ground-truth transcripts, so the authors place most of the gap in rendering text rather than reading it, and say a stronger MLLM alone would bring limited gains.
The repository’s code and the ReasonText annotations are licensed CC BY-NC 4.0, so they may be shared and adapted only for non-commercial purposes and with attribution, while model weights and the source datasets follow their own terms, and the bundled DiT4SR backbone is redistributed under its own licence (GitHub).
Risks of inferred text
The authors say that text recovered through reasoning is a plausible inference, not evidence, and that it can sometimes differ from what the image truly contained. They advise against treating such outputs as ground truth in forensic analysis or document verification, and urge caution with images that show private information, for example license plates or personal documents. They add that the decoupled design helps: the MLLM caption lays the inferred text open to inspection, so an error can be traced to the component that caused it. The authors note that the decoupled design makes failures traceable, because the caption shows the inferred text and lets one check whether the MLLM misread it or the restoration model failed to render it.
How we built this
Published resultsTable F of the paper: Reasoning Transfer via Captioning / ReasonText benchmark
Rows of Table F of the paper (arXiv HTML version), copied cell for cell by our code; a two-row header is shown as "group: column". These are the authors' own figures, not an independent test. The table has 12 columns; 5 shown, with every column that holds a figure this article gives or a measure it names. Not shown: Fidelity (2 columns), Perceptual (5 columns). Rows: 17 of 25, with every row whose value this article gives, or that it names; not shown: rows 14-21; whole blocks only (a row with the rows "+ …" under it). Of its 4 blocks, not shown: C (4 rows); shown in part: E (4 of 8 rows).
Sources
The sources this article was written from. Links open the original page.
- arXivReading the Unreadable: Text-Aware Image Super-Resolution Needs Reasoningarxiv.org
- arXivReading the Unreadable: Text-Aware Image Super-Resolution Needs Reasoningarxiv.org
- Hugging FaceLicence declared in the front matter of the dataset card Jasonleex1995/ReasonText on Hugging Face, linked by https://jasonleex1995.github.io/TAISR_needs_Reasoninghuggingface.co
- GitHubLicence or terms of use linked by https://github.com/jasonleex1995/TAISR_needs_Reasoningraw.githubusercontent.com