Reading the Unreadable

1. Benchmark

To answer this question, we construct the ReasonText benchmark, the first TAISR benchmark that separates locally readable text from reasoning-required text.

We label each text instance by how a human can read it in the LR image:

Level 1 (locally readable): a human can read the text given only the cropped LR text image.

Level 2 (reasoning-required): a human can only read it from the LR image by reasoning over the surrounding context.

2. Diagnosis

A: GAN-based · B: Multi-step diffusion · C: Few-step diffusion · D: AR-based · E: Text-aware SR

Group	Model	Fidelity		Perceptual					Text Restoration
		Fidelity		Reference		No Reference			Text Restoration
		PSNR↑	SSIM↑	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
A	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
	SUPIR	24.2971	0.7098	0.2622	0.1947	0.4608	63.5998	0.5922	55.2984	68.5291	22.5670
	FaithDiff	24.2038	0.7363	0.2061	0.1623	0.5095	71.6540	0.5821	52.0097	65.1653	19.4640
	DiT4SR	22.6539	0.6933	0.2477	0.1830	0.4973	70.8502	0.6074	59.1555	73.1471	24.5416
C	ResShift	24.9170	0.7645	0.2219	0.1839	0.4110	63.3033	0.4902	47.7060	61.0034	14.8096
	SinSR	25.1644	0.7668	0.2116	0.1747	0.4551	66.1994	0.5574	43.1181	55.9863	11.2835
	OSEDiff	23.9596	0.7533	0.2147	0.1666	0.4803	71.1559	0.5672	44.6204	58.4379	10.4372
D	PURE	22.2970	0.6706	0.2652	0.1893	0.5233	70.0036	0.6531	34.3890	45.7241	6.3470
E	TeReDiff	23.5188	0.7233	0.2362	0.1796	0.5638	70.5181	0.6183	51.2383	65.1083	16.9252
E	UniT	22.7672	0.6503	0.2814	0.2072	0.5159	68.7382	0.6315	61.3886	76.4538	24.1185

For each metric, the best score among the twelve SR models is shown in bold (the HR oracle is excluded).

We evaluate twelve recent SR models on ReasonText, including real-world SR models for natural images (A–D) and TAISR models designed for text (E).

All models struggle to restore reasoning-required text, regardless of architecture, generative backbone, or training paradigm, and even TAISR models designed specifically for text are no exception.

These results answer our question:

TAISR does need reasoning.

3. Method

The remaining question is how to supply this reasoning to SR models. Many recent SR models already take a text caption as an additional input, and our method builds on two observations:

Observation 1. These text-conditioned SR models restore reasoning-required text substantially better when the caption contains the ground-truth transcripts.

Observation 2. Modern multimodal large language models (MLLMs) can read much of the reasoning-required text in an LR image by reasoning over its context.

RTC method: Caption by Reasoning then Restore by Conditioning — Overall pipeline of Reasoning Transfer via Captioning (RTC), illustrated with an example.

Inspired by these two observations, we propose Reasoning Transfer via Captioning (RTC), which works in two stages:

Stage 1 (Caption by Reasoning): an MLLM reads the LR image with a reasoning prompt and produces a two-sentence caption: a scene description, followed by the text it infers by reasoning.

Stage 2 (Restore by Conditioning): this caption replaces the default caption of a text-conditioned SR model, which restores the image.

Both models stay frozen, so RTC can be attached to any text-conditioned SR model.

Group	Model	Fidelity		Perceptual					Text Restoration
		Fidelity		Reference		No Reference			Text Restoration
		PSNR↑	SSIM↑	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
	+ RTC	24.2724	0.7387	0.2020	0.1602	0.4975	71.4577	0.5797	61.0231	74.0023	28.9140
	DiT4SR	22.6539	0.6933	0.2477	0.1830	0.4973	70.8502	0.6074	59.1555	73.1471	24.5416
	+ RTC	22.6816	0.6936	0.2479	0.1830	0.5009	70.6745	0.6110	73.5688	84.7777	45.8392
C	OSEDiff	23.9596	0.7533	0.2147	0.1666	0.4803	71.1559	0.5672	44.6204	58.4379	10.4372
C	+ RTC	23.8903	0.7504	0.2171	0.1682	0.4886	71.2281	0.5767	49.4113	64.2531	12.6939
E	TeReDiff	23.5188	0.7233	0.2362	0.1796	0.5638	70.5181	0.6183	51.2383	65.1083	16.9252
	+ RTC	23.3884	0.7212	0.2377	0.1795	0.5777	70.9852	0.6237	56.6382	70.1254	23.2722
	UniT	22.7672	0.6503	0.2814	0.2072	0.5159	68.7382	0.6315	61.3886	76.4538	24.1185
	+ RTC	22.8425	0.6555	0.2751	0.2041	0.5169	68.8130	0.6338	70.9704	82.5542	42.3131

For each + RTC value, blue/red marks an improvement/drop over the baseline.

RTC consistently improves all six text-conditioned SR models, especially on reasoning-required text, while preserving general image quality.

Remarkably, with RTC, the real-world SR model DiT4SR outperforms UniT, a TAISR model that builds on the same backbone and adds text-aware modules. This suggests that, once reasoning is supplied, current text-aware modules may bring no additional benefit.

It's time to design TAISR that reasons!