Paper deep dive
TOLA: Text-aware One-Step Latent Adaptation for Diffusion-based Text Image Super-Resolution
Yike Xu, Yue Shi, Yong Guo, Jiezhang Cao
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 9/25/2026, 5:09:14 AM
Summary
The paper introduces TOLA, a Text-aware One-step Latent Adaptation framework for Text Image Super-Resolution (TSR). TOLA addresses the high computational cost and error amplification issues of existing multi-step diffusion-based methods by utilizing a single-step inference process. It employs a confidence-weighted text conditioning module to suppress unreliable OCR predictions and a lightweight latent residual correction module to fix structured errors in the clean latent space. TOLA achieves state-of-the-art performance on CTR-TSR-Test and RealCE-200 benchmarks, significantly outperforming DiffTSR in PSNR and inference speed.
Entities (12)
Relation Signals (10)
TOLA → solves → Text Image Super-Resolution
confidence 95% · we propose TOLA, a Text-aware One-step Latent Adaptation framework... Text image super-resolution (TSR) aims to recover visually faithful and readable text
TOLA → outperforms → DiffTSR
confidence 92% · our TOLA consistently surpasses existing diffusion-based TSR methods... TOLA outperforms the official 200-step DiffTSR in all nine metrics
TOLA → achieveshigherscorein → PSNR
confidence 90% · our TOLA consistently surpasses existing diffusion-based TSR methods by at least 2.72 dB in PSNR
TOLA → achieveslowerscorein → FID
confidence 90% · reducing FID by 12.2271.
TOLA → evaluatedon → CTR-TSR-Test
confidence 90% · Extensive experiments demonstrate our state-of-the-art performance... on both CTR-TSR-Test
TOLA → evaluatedon → RealCE-200
confidence 90% · Extensive experiments demonstrate our state-of-the-art performance... on both CTR-TSR-Test and RealCE-200 benchmarks.
TOLA → uses → TransOCR
confidence 88% · the frozen TransOCR instance used for conditioning predicts a token sequence
TOLA → uses → IDM
confidence 85% · the LoRA-adapted IDM performs one noise prediction
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract: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 ($\times 4$) 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.
Tags
Links
- Source: https://arxiv.org/abs/2609.29240v1
- Canonical: https://arxiv.org/abs/2609.29240v1
Trouble viewing inline? Open PDF directly →
Full Text
69,568 characters extracted from source content.
Expand or collapse full text
TOLA: Text-aware One-Step Latent Adaptation for Diffusion-based Text Image Super-Resolution Yike Xu †thanks: Work done during an internship at Shanghai Jiao Tong University. Affiliation: Shanghai Jiao Tong University Affiliation: University of Chinese Academy of Sciences Yue Shi Affiliation: Shanghai Jiao Tong University Affiliation: Shanghai AI Laboratory Yong Guo Affiliation: Shanghai Jiao Tong University Jiezhang Cao †thanks: Corresponding author. Affiliation: Shanghai Jiao Tong University Abstract 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. Introduction Text image super-resolution (TSR) aims to recover high-resolution text images from degraded low-resolution observations while preserving both visual details and character-level semantics (Zhao et al., 2021). This challenge becomes more pronounced in the blind setting, where the degradation process is unknown. Unlike natural-image super-resolution, TSR must preserve not only perceptual quality but also character-level semantic accuracy, as a single missing, merged, or hallucinated stroke may lead to an incorrect transcription even when the reconstructed image appears visually sharp. Therefore, TSR models need to recover fine-grained visual structures while maintaining semantic fidelity. Early TSR methods enhance reconstruction by incorporating sequence modeling, text-aware attention, recognition guidance, and character-structure priors (Wang et al., 2020; Ma et al., 2022; Ma et al., 2023a; Li et al., 2023). These methods perform well when the low-resolution input still contains reliable structural or semantic cues. Under severe degradations, however, missing and merged strokes make both visual reconstruction and text recognition highly ambiguous. Diffusion models have recently emerged as a promising solution because their generative priors can synthesize plausible high-frequency details that deterministic restoration networks often fail to recover (Ho et al., 2020; Rombach et al., 2022; Kawar et al., 2022; Saharia et al., 2023; Wang et al., 2024a). Existing diffusion-based TSR methods typically rely on multi-step prediction of either the high-resolution image or its text-aware prior (Zhang et al., 2024; Xu et al., 2026), as shown in figure 1. Although iterative refinement improves generative capacity, it also introduces two fundamental limitations. First, repeatedly evaluating image and text networks incurs computational cost and inference latency. Second, the text prior used to guide reconstruction is itself inferred from the degraded observation and may therefore be incorrect (Yang et al., 2024). When an erroneous prior is repeatedly injected into the denoising process, image and text predictions may reinforce each other, progressively amplifying an early recognition error into a sharp yet semantically incorrect character. Figure 1: Comparison of three different diffusion-based TSR inference paradigms. (a) DiffTSR performs 200-step iterative text-prior and image restoration. (b) PRISM uses one-step image reconstruction with 16-step text-prior recovery. (c) TOLA constructs the text condition once, evaluates IDM once, and applies a Latent Residual Correction (LRC) for clean-latent correction. Prior iterative text-image refinement may repeatedly propagate inaccurate semantic cues through multiple denoising stages, leading to accumulated recognition errors. In contrast, TOLA avoids error amplification by designing a one-step text-conditioned IDM. Directly reducing multi-step diffusion-based TSR to a single image-denoising step is non-trivial due to two challenges. First, reliable text guidance must be constructed in a single pass without repeatedly injecting uncertain OCR predictions into image reconstruction. Second, the clean latent estimated from one high-noise prediction inevitably contains structured errors, especially around thin strokes, character boundaries, and ambiguous local patterns, which can no longer be corrected by subsequent reverse updates. Introducing another iterative prior-recovery or refinement process would merely shift, rather than eliminate, the computational burden. The key is therefore to combine one-time semantic conditioning with explicit, non-iterative latent residual correction. Based on this observation, we propose TOLA, a Text-aware One-step Latent Adaptation framework. Our central insight is that semantic conditioning and latent residual correction need not be coupled along the same reverse trajectory. Instead, reliable semantic evidence can be selected once before image prediction, while structured reconstruction errors can be modeled explicitly afterward in the clean-latent space. As illustrated in figure 2, TOLA consists of two complementary modules. First, the confidence-weighted text conditioning module uses frozen TransOCR to predict text tokens and token-wise confidence scores, then constructs the semantic condition once via the frozen MoM module. By suppressing uncertain OCR predictions, this module reduces the risk that an unreliable text hypothesis will dominate image reconstruction. Second, the lightweight latent residual correction module explicitly estimates the structured residual remaining in the initial clean latent and restores missing or distorted stroke details without introducing another reverse trajectory. Our main contributions are summarized as follows: • We propose TOLA, which reformulates the original multi-step image-text diffusion pipeline into a non-iterative framework with one-shot text conditioning and one-step latent prediction. • We introduce confidence-weighted text conditioning, which constructs the semantic condition once and suppresses unreliable character priors using OCR confidence scores. This prevents erroneous text information from being repeatedly injected and amplified during denoising, yielding reliability-aware semantic guidance. • Extensive experiments demonstrate that TOLA achieves state-of-the-art performance across all reported metrics on CTR-TSR-Test (×4× 4) and RealCE-200. TOLA outperforms the official 200-step DiffTSR in all nine metrics, improving PSNR by 4.2137 dB while reducing FID by 12.2271. Meanwhile, it shortens the average inference time from 9.9548s to 0.1319s, achieving a 75.5× speedup. Related Work Text-Aware Scene Text Image Super-Resolution. Generic SR optimizes visual fidelity without preserving character identity (Dong et al., 2016; Wang et al., 2018; Chen et al., 2022b; Liang et al., 2021). TSR models strokes, sequences, layout, deformation, and location (Wang et al., 2020; Chen et al., 2021; Ma et al., 2022; Chen et al., 2022a; Zhao et al., 2022; Zhu et al., 2023b; Guo et al., 2023; Wei et al., 2025; TomyEnrique et al., 2024). Recognition, style, and structure priors provide guidance (Ma et al., 2023a; Li et al., 2023; Zhu et al., 2023a; Park and Ko, 2025; Yuan et al., 2025), but cues extracted from degraded inputs remain unreliable (Kong et al., 2024). Diffusion-Based Text Image Restoration. Diffusion TSR conditions denoising on masks, LR text, recognition, or character embeddings (Liu et al., 2025; Zhou et al., 2024; Noguchi et al., 2024; Singh et al., 2024). DiffTSR couples IDM and TDM through MoM; Boosting DiffTSR adds mixed training, ResShift, cross-attention, and confidence weighting (Zhang et al., 2024; Pan et al., 2025). TextSR, TEXTS-Diff, and TADiSR address multilingual or real-world restoration (Ye et al., 2025; He et al., 2026; Hu et al., 2025); DualTSR and TeReDiff combine text diffusion or recognition with restoration (Niu et al., 2026; Min et al., 2026). These methods retain iterative refinement. One-Step Diffusion Restoration. DDIM, ResShift, SinSR, and AddSR shorten or distill diffusion trajectories (Song et al., 2021; Yue et al., 2023; Wang et al., 2024b; Tai et al., 2026; Dong et al., 2025); OSEDiff uses LoRA and score distillation (Hu et al., 2022; Wu et al., 2024), while FiDeSR combines frequency injection with latent residual refinement (Kim et al., 2026). These methods target generic SR rather than the coupled DiffTSR process. PRISM is text-specific but uses K=16K=16 Euler steps for prior recovery (Xu et al., 2026). TOLA instead adapts pretrained DiffTSR to one IDM evaluation by constructing confidence-weighted TransOCR–MoM conditioning once and correcting the latent after noise-to-clean conversion. Method Figure 2: TOLA architecture. The LR image is encoded and q-sampled at fixed t=999t=999, while confidence-weighted TransOCR tokens and latent features form one MoM condition. A LoRA-adapted IDM predicts noise once, and LRC corrects the clean latent before frozen VAE decoding. Snowflakes and flames denote frozen and trainable components, respectively. Overview As shown in figure 2, TOLA converts iterative DiffTSR into a single-pass restoration path. TransOCR predicts tokens and confidence scores, while the LoRA-adapted VAE encoder maps the LR image to LR z_LR and q-sampling obtains tz_t at t=999t=999. MoM fuses confidence-weighted token embeddings with [LR,t][ z_LR,z_t] to construct condC_cond once. Conditioned on condC_cond, the LoRA-adapted IDM performs one noise prediction, which is converted into the initial clean latent ^0 z_0. LRC then corrects the residual LR−^0 z_LR- z_0 before VAE decoding. Only LoRA and LRC are optimized, and inference requires neither TDM nor iterative reverse updates. Text-Aware One-Step Adaptation Fixed-Timestep q-Sampling. The LoRA-adapted VAE encoder first encodes the degraded input image LR x_LR into an unscaled latent representation LR z_LR, LR=ℰVAE(LR). z_LR=E_VAE( x_LR). (1) Following DiffTSR q-sampling, we sample a noisy latent at the fixed, zero-indexed timestep t=999t=999: t=α¯tLR+1−α¯tϵ,ϵ∼(,),z_t= α_t\, z_LR+ 1- α_t\, , ( 0, I), (2) where α¯t=∏τ=0t(1−βτ) α_t= _τ=0^t(1- _τ) follows the 1000-step linear schedule, with β0=0.0015 _0=0.0015 and β999=0.0205 _999=0.0205. We draw one Gaussian sample per input and supply LR z_LR and tz_t to IDM on the VAE latent scale. Unlike TOLA, OSEDiff feeds the unnoised LR latent to its one-step U-Net (Wu et al., 2024). Confidence-Weighted Text Conditioning. The frozen TransOCR instance used for conditioning predicts a token sequence and the corresponding token-wise confidence scores, (~nat,) ( c^nat, α) =TransOCR(LR), = ( x_LR), (3) ~ c =Map(Norm(~nat)). = \! ( ( c^nat) ). Norm normalizes characters to Simplified Chinese; Map maps them to the IDM vocabulary with α alignment. Following Boosting DiffTSR (Pan et al., 2025), confidence-weighted embeddings enter MoM, which computes [cond,cond]=ℳ(s[LR,t],⊙Emb(~),t),[I_cond,C_cond]=M\! (s\,[ z_LR,z_t], α ( c),t ), (4) where s=0.18215s=0.18215 scales image latents and ⊙ broadcasts token confidences over embeddings. IDM uses only condC_cond for cross-attention; condI_cond is discarded. Confidence is neither calibrated nor thresholded, and MoM runs once. One-Step Clean-Latent Reconstruction. The LoRA-adapted IDM receives channel-wise concatenated noisy and LR latents and predicts ϵ^=f([t,LR],t,cond), _ θ=f_ θ ([z_t, z_LR],t,C_cond ), (5) where [⋅,⋅][·,·] is channel-wise concatenation and θ combines frozen base and trainable LoRA parameters. Since DiffTSR predicts noise, the clean latent is ^0=t−1−α¯tϵ^α¯t. z_0= z_t- 1- α_t\, _ θ α_t. (6) No reverse update follows equation 6; ^0 z_0 is passed directly to LRC. Algorithm 1 summarizes inference. Algorithm 1 TOLA one-step inference. 0: LR image LR x_LR; fixed t=999t=999 0: Restored image SRx_SR 1: LR←ℰVAE(LR) z_LR←E_VAE( x_LR); (~nat,)←TransOCR(LR)( c^nat, α)← ( x_LR) 2: ~←Map(Norm(~nat)) c← ( ( c^nat)) 3: Draw ϵ∼(,) ( 0, I); construct tz_t with equation 2 4: [cond,cond]←ℳ(s[LR,t],⊙Emb(~),t)[I_cond,C_cond]←M(s\,[ z_LR,z_t], α ( c),t) 5: ϵ^←f([t,LR],t,cond) _ θ← f_ θ([z_t, z_LR],t,C_cond) 6: Reconstruct ^0 z_0 with equation 6 7: ←LR−^0r← z_LR- z_0; Δ←gϕ([LR,]) r← g_ φ([ z_LR,r]) 8: ^0corr←LR−(+Δ) z_0^corr← z_LR-(r+ r); SR←VAE(s−1^0corr)x_SR←D_VAE(s^-1 z_0^corr) TOLA evaluates the VAE encoder/decoder, TransOCR, MoM, IDM, and LRC once each while omitting TDM; DiffTSR evaluates IDM and TDM 200 times each. Method PSNR↑ SSIM↑ LPIPS↓ DISTS↓ FID↓ T-ACC↑ T-NED↑ P-ACC↑ P-NED↑ SRCNN 22.0288 0.5860 0.6021 0.3886 156.9406 0.4255 0.6107 0.3303 0.5279 ESRGAN 21.7741 0.5588 0.6455 0.3979 155.7055 0.4188 0.6061 0.3287 0.5273 NAFNet 21.5568 0.5678 0.5924 0.3776 144.9606 0.4187 0.5998 0.2998 0.4870 TSRN 20.4381 0.5840 0.6768 0.3926 140.2070 0.3307 0.5192 0.2238 0.4196 TBSRN 20.6883 0.5840 0.6510 0.3918 137.1033 0.3539 0.5425 0.2574 0.4563 TATT 20.9523 0.5846 0.6614 0.3896 125.6047 0.3558 0.5450 0.2491 0.4466 MARCONet 21.3460 0.6132 0.4386 0.3566 102.7384 0.3895 0.5735 0.2796 0.4600 PRISM 22.08-- – 0.2314 – 12.57-- 0.4212 0.6644 – – DiffTSR 20.5958 0.6261 0.3103 0.2428 24.9120 0.4645 0.6519 0.3970 0.5873 TeReDiff 19.6333 0.5125 0.4714 0.3466 77.0387 0.2893 0.4491 0.1972 0.3592 Ours 24.8095 0.7235 0.2288 0.1953 12.6849 0.5401 0.7370 0.4626 0.6585 Table 1: Results on 8,089 CTR-TSR-Test ×4× 4 images. Non-diffusion methods are retrained on the CTR-TSR split, and all unified-protocol outputs share one evaluation implementation. T/P denote TransOCR/PaddleOCR; bold marks the best unified-protocol result. PRISM (Xu et al., 2026) is a BTL-test ×4× 4/P-OCRv5 cross-protocol reference; dashes preserve its reported precision. Note: “–” denotes unavailable values. PRISM results are from its arXiv paper due to unavailable code. Latent Residual Correction A FiDeSR-inspired LRC corrects local errors after noise-to-clean conversion (Kim et al., 2026). FiDeSR refines a directly predicted residual, whereas TOLA corrects the clean-latent residual without changing semantic conditioning. Given the initial clean-latent estimate ^0 z_0, we define =LR−^0.r= z_LR- z_0. (7) From the concatenated input [LR,][ z_LR,r], LRC predicts Δ=gϕ([LR,]), r=g_ φ([ z_LR,r]), (8) The trainable gϕg_ φ uses a 1×11× 1 projection from 6 to 32 channels, one RRDB of three five-layer RDBs with 3×33× 3 convolutions, and a 1×11× 1 projection to 3 channels, with growth width 16 and residual scaling 0.2. The corrected residual and latent are corr=+Δ,^0corr=LR−corr.r^corr=r+ r, z_0^corr= z_LR-r^corr. (9) The frozen VAE decoder then reconstructs SR=VAE(s−1^0corr),x_SR=D_VAE\! (s^-1 z_0^corr ), (10) The inherited scale is s=0.18215s=0.18215. Zero-initializing the LRC output projection initially preserves ^0 z_0 and learns only the residual correction. Training Objective TOLA uses pixel, perceptual, and recognition objectives: ℒ= = ℒ1(SR,HR)+ℒLPIPSVGG(SR,HR) _1(x_SR, x_HR)+L_LPIPS^VGG(x_SR, x_HR) (11) +0.02ℒOCRCE(OCR(SR),GT). +0.02\,L_OCR^CE\! ( (x_SR), y_GT ). HR x_HR and GT y_GT are the ground-truth image and transcription; ℒ1L_1 is ℓ1 _1, ℒLPIPSVGGL_LPIPS^VGG is LPIPS-VGG, and ℒOCRCEL_OCR^CE is token-level cross-entropy. Two frozen TransOCR instances provide conditioning and OCR supervision; OCR gradients update LoRA and LRC. Conditioning uses the IDM vocabulary, whereas ℒOCRCEL_OCR^CE retains the native TransOCR labels. No latent ℓ1 _1, noise-prediction, distillation, or teacher loss is used. Experiments Experimental Setup Datasets. We use 62,371 training and 1,273 validation images from the CTR-TSR preprocessing set (Zhang et al., 2024). LR inputs are generated online with Real-ESRGAN or BSRGAN degradations at scales 1,2,4\1,2,4\ (Wang et al., 2021; Zhang et al., 2021); all images use a 128×512128× 512 canvas. Evaluation uses all 8,089 pairs in CTR-TSR-Test ×4× 4. For real-image generalization, we use two complementary test sets. RealCE-200 contains 200 filtered and deduplicated pairs from the RealCE validation split (Ma et al., 2023b), using matched official 13m13\,m and 52m52\,m crops as LR and reference images. We construct RT50 to extend text-category and acquisition-condition coverage beyond existing paired benchmarks. Its 50 images, collected through multiple routes, include 17 Chinese, 17 English, and 16 numeric samples spanning varied blur, illumination, viewpoints, resolutions, and backgrounds. A predefined output-independent protocol screens content validity, geometry, degradation difficulty, quality, and duplication. Neither set is used for training, validation, or checkpoint selection. Evaluation Protocol and Metrics. To ensure a fair comparison, we adapt all non-diffusion methods to ×4× 4 restoration and retrain them on the same CTR-TSR training split. All unified-protocol methods are evaluated on the same 8,089-pair manifest at 128×512128× 512 using identical filename matching and metric implementations. We report luminance-channel PSNR and SSIM without border cropping (Wang et al., 2004), LPIPS-AlexNet (Zhang et al., 2018), DISTS (Ding et al., 2022), and FID (Heusel et al., 2017); training instead uses LPIPS-VGG. TransOCR predictions undergo full-to-half-width and simplified-Chinese conversion and whitespace removal while preserving case. ACC denotes exact-match accuracy, and NED denotes mean normalized edit-distance similarity. Independent PaddleOCR ACC and NED (Du et al., 2020) provide a recognizer cross-check. Compared Methods. We compare TOLA with SRCNN, ESRGAN, NAFNet, TSRN, TBSRN, TATT, MARCONet, DiffTSR, and TeReDiff, and additionally include the published PRISM results as a cross-protocol reference. Implementation Details. Starting from the official CTR-trained DiffTSR checkpoint, we optimize rank-4 LoRA and LRC for 100K steps on four NVIDIA RTX PRO 6000 GPUs with batch size 16 per GPU. FP32 AdamW uses learning rate 5×10−55× 10^-5, (β1,β2)=(0.9,0.999)( _1, _2)=(0.9,0.999), weight decay 0.01, 500-step linear warm-up, and gradient clipping at 1.0. Training takes about 55 hours, and validation selects the checkpoint. Loss weights are 1, 1, and 0.02. Each image is restored once without sample selection or averaging. Figure 3: Comparison on five selected CTR-TSR-Test ×4× 4 images covering Chinese, English, and numeric text. Each column presents an aligned test sample, and the rows show the LR input, displayed baselines, Ours, and the HR reference. Diagnostic Protocol. All trainable variants in Table 4 share CTR data, DiffTSR initialization, data order, optimization, and the 8,089-image evaluation. Conditioning controls alter only the fixed checkpoint input; GT prior is an oracle. Paired comparisons use identical per-image noise. Main Results Quantitative Comparisons. Table 1 shows that TOLA surpasses DiffTSR in all nine metrics, gaining 4.2137 dB PSNR and 0.0756 TransOCR ACC while reducing FID by 12.2271. PaddleOCR confirms this trend: ACC/NED rise from 0.3970/0.5873 to 0.4626/0.6585. CTR-TSR Qualitative Results. Figure 3 shows more coherent strokes and closer character shapes and layouts across Chinese, English, and numeric examples. Cross-Dataset Results. Table 2 shows TOLA leading all six metrics, exceeding DiffTSR by 1.5953 dB PSNR, 4.50 points TransOCR ACC, and 8.00 points PaddleOCR ACC. Method PSNR↑ SSIM↑ T-ACC↑ T-NED↑ P-ACC↑ P-NED↑ SRCNN 19.1609 0.6024 0.3850 0.5837 0.4950 0.6670 ESRGAN 19.2062 0.5971 0.3850 0.5799 0.5000 0.6710 NAFNet 18.9038 0.5920 0.3650 0.5704 0.4750 0.6406 TSRN 18.6767 0.6009 0.2850 0.5103 0.3850 0.5649 TBSRN 18.9903 0.5947 0.3400 0.5364 0.3950 0.5980 TATT 18.9725 0.5954 0.3300 0.5291 0.4000 0.5970 MARCONet 18.3506 0.6198 0.3500 0.5408 0.3800 0.5972 DiffTSR 17.6287 0.5970 0.3500 0.5402 0.4300 0.6188 TeReDiff 17.8180 0.5272 0.2900 0.4826 0.3350 0.5508 Ours 19.2240 0.6199 0.3950 0.5841 0.5100 0.6720 Table 2: RealCE-200 cross-dataset results on the same 200 pairs. RealCE is excluded from training and checkpoint selection; T/P denote TransOCR/PaddleOCR. Evaluation on RT50. Figure 4 shows that TOLA better preserves strokes and character shapes on Chinese, English, and numeric text under varied real degradations. Figure 4: RT50 comparison on Chinese, English, and numeric text under varied real degradations. Rows show LR inputs, displayed baselines, and Ours; examples cover blur, illumination, viewpoint, resolution, and background variations. Figure 5: Quality–latency comparison on CTR-TSR-Test ×4× 4. Panels (a) and (b) plot PSNR and TransOCR ACC against synchronized batch-1 inference time measured on 100 fixed CTR inputs using one NVIDIA RTX PRO 6000. Inference Cost and Latency. We use synchronized batch-1 timing on 100 fixed CTR images with one NVIDIA RTX PRO 6000, excluding loading. Table 3 reports complete-trajectory MACs at 128×512128× 512 and restoration-module parameters excluding VAEs and auxiliary text modules. TOLA uses 297.497 G MACs and 0.1319 seconds per image, versus 59,213.970 G and 9.9548 seconds for DiffTSR, while its optimized and saved state is 2.46M parameters. Figure 5 visualizes the quality–latency trade-off. Method # Params. (M) MACs (G) Inference Time (s) SRCNN 0.057 3.748 0.0119 ESRGAN 16.70 73.453 0.0335 NAFNet 67.89 63.195 0.0324 TSRN 2.68 0.904 0.0148 TBSRN 3.21 2.535 0.0209 TATT 7.61 1.269 0.0287 MARCONet 87.90 466.178 0.1986 DiffTSR 874.00 59,213.970 9.9548 TeReDiff 1682.54 26,839.909 3.8499 Ours 876.23 297.497 0.1319 Table 3: Restoration parameters, MACs, and batch-1 inference time on 100 CTR images. VAEs and text modules are excluded; TOLA stores 2.46M adapted parameters. Ablations and Diagnostic Analyses Core Components. Table 4 shows that LoRA drives one-step adaptation and LRC adds 0.3847 dB PSNR while reducing LPIPS/DISTS by 0.0121/0.0017; its ACC change is not significant (p=0.584p=0.584). A parameter-matched plain correction loses 0.8249 dB PSNR and increases FID by 4.2491, so capacity alone does not explain the gain. IDM-only LoRA remains much closer to Full than VAE-only LoRA. Removing MoM or bypassing its latent fusion degrades every metric. Relative to uniform confidence, confidence weighting adds 0.1908 dB PSNR and 0.0134/0.0138 TransOCR ACC/NED while reducing LPIPS/DISTS/FID by 0.0213/0.0160/4.2766. Null context confirms the value of predicted text, and the GT oracle adds another 0.0862 ACC and 0.0716 NED, revealing headroom from better priors. Figure 6: Core component visualization. (a) LoRA enables one-step adaptation and LRC corrects local structure over frozen and single-module controls. (b) No MoM and direct-token conditioning cause character errors; LR and HR provide references. Configuration PSNR↑ SSIM↑ LPIPS↓ DISTS↓ FID↓ T-ACC↑ T-NED↑ P-ACC↑ P-NED↑ Frozen one-step 20.0416 0.5342 0.5254 0.3464 119.4517 0.2276 0.3925 0.1278 0.2918 LRC only 21.9696 0.6021 0.4880 0.3301 110.0081 0.3785 0.5713 0.2440 0.4337 LoRA only 24.4220 0.7109 0.2410 0.1970 12.7420 0.5378 0.7337 0.4570 0.6564 Full 24.8095 0.7235 0.2288 0.1953 12.6849 0.5401 0.7370 0.4626 0.6585 Plain residual correction 23.9846 0.6991 0.2843 0.2180 16.9340 0.5080 0.7047 0.4256 0.6243 No MoM 23.9624 0.6984 0.2833 0.2170 16.1812 0.5060 0.7049 0.4117 0.6197 Direct token condition 24.0026 0.6993 0.2846 0.2186 16.7650 0.5079 0.7038 0.4093 0.6185 IDM LoRA only 23.8812 0.6987 0.2844 0.2174 16.7593 0.5104 0.7058 0.4232 0.6254 VAE LoRA only 22.4861 0.6141 0.4635 0.3221 108.0600 0.3989 0.5926 0.2597 0.4505 Null context 24.5495 0.7118 0.2695 0.2239 22.1417 0.5308 0.7280 0.4390 0.6438 Uniform confidence 24.6159 0.7149 0.2502 0.2113 16.9615 0.5258 0.7224 0.4285 0.6356 GT prior (oracle) 24.9148 0.7283 0.2216 0.1920 12.7565 0.6253 0.8079 0.5520 0.7387 Table 4: Ablation on CTR-TSR-Test ×4× 4. Full uses IDM/VAE LoRA, LRC, and confidence-aware MoM. Plain is parameter-matched; No MoM removes the text branch; direct tokens bypass fusion; GT is oracle; T/P denote TransOCR/PaddleOCR. Objective and Hyperparameter Diagnostics. Rank 4 outperforms ranks 2 and 8. Small, default, and large LRCs use (hidden,growth,RRDBs)=(16,8,1)(hidden,growth,RRDBs)=(16,8,1), (32,16,1)(32,16,1), and (32,16,2)(32,16,2) with 0.045235M, 0.180323M, and 0.360323M parameters; the default gives the best trade-off. Removing ℓ1 _1 most reduces PSNR, removing LPIPS most harms perceptual metrics, and OCR weight 0 lowers recognition; 0.05 brings no gain. Among tested timesteps, t=999t=999 performs best. Text-Prior Reliability and Error Analysis. Figure 7 shows TOLA outperforms DiffTSR across confidence bins and text categories, reducing substitutions, deletions, and insertions. The exact-match gap between correct and incorrect priors identifies recognition errors as the main limitation. Figure 7: Recognition diagnostics on CTR-TSR-Test ×4× 4: (a) ACC across confidence bins, (b) ACC across text categories, and (c) character-level edit counts. Inference Controls and Stochastic Stability. Controls favor t=999t=999; zero noise and disabled VAE LoRA reduce quality, while LR-input and frozen-VAE controls rule out copying or VAE-only reconstruction. Across three draws, maximum standard deviations are 0.0019 dB PSNR and 0.0012 ACC. Paired Uncertainty. Paired bootstrap intervals exclude zero for LRC effects on PSNR, SSIM, LPIPS, and DISTS, but not NED; exact McNemar testing finds no significant ACC change. Predicted conditioning improves ACC over null context, and the GT-prior gap confirms prior errors. Conclusion We presented TOLA, a text-aware one-step adaptation of DiffTSR. TOLA constructs a confidence-aware TransOCR–MoM condition once, adapts IDM and the VAE encoder with rank-4 LoRA, and corrects clean-latent residual errors with LRRB without invoking TDM. It improves all nine CTR-TSR-Test metrics over DiffTSR while reducing IDM/TDM evaluations from 200/200 to 1/0 and latency from 9.9548 to 0.1319 seconds per image. The optimized and saved state contains only 2.46M parameters, while inference retains the pretrained backbone without introducing another iterative restoration trajectory. Component ablations and independent-OCR evaluation support the complementary roles of LoRA, LRRB, and confidence-aware conditioning. Results on RealCE-200 and RT50 further demonstrate generalization across paired camera observations, Chinese, English, and numeric text, and varied real capture conditions. These results establish efficient one-step diffusion TSR without iterative image–text refinement. References Chen et al. (2021) J. Chen, B. Li, and X. Xue Scene Text Telescope: text-focused scene image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 12026–12035. Cited by: Related Work. Chen et al. (2022a) J. Chen, H. Yu, J. Ma, B. Li, and X. Xue Text Gestalt: stroke-aware scene text image super-resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 36, p. 285–293. Cited by: Related Work. Chen et al. (2022b) L. Chen, X. Chu, X. Zhang, and J. Sun Simple baselines for image restoration. In European Conference on Computer Vision, p. 17–33. Cited by: Related Work. Ding et al. (2022) K. Ding, K. Ma, S. Wang, and E. P. Simoncelli Image quality assessment: unifying structure and texture similarity. IEEE Transactions on Pattern Analysis and Machine Intelligence 44 (5), p. 2567–2581. Cited by: Evaluation Protocol and Metrics.. Dong et al. (2016) C. Dong, C. C. Loy, K. He, and X. Tang Image super-resolution using deep convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 38 (2), p. 295–307. Cited by: Related Work. Dong et al. (2025) L. Dong, Q. Fan, Y. Guo, Z. Wang, Q. Zhang, J. Chen, Y. Luo, and C. Zou TSD-SR: one-step diffusion with target score distillation for real-world image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 23174–23184. Cited by: Related Work. Du et al. (2020) Y. Du, C. Li, R. Guo, X. Yin, W. Liu, J. Zhou, Y. Bai, Z. Yu, Y. Yang, Q. Dang, and H. Wang P-OCR: a practical ultra lightweight OCR system. External Links: 2009.09941 Cited by: Evaluation Protocol and Metrics.. Guo et al. (2023) H. Guo, T. Dai, G. Meng, and S. Xia Towards robust scene text image super-resolution via explicit location enhancement. In International Joint Conference on Artificial Intelligence, p. 782–790. Cited by: Related Work. He et al. (2026) H. He, X. Zhan, Y. Bai, R. Lan, L. Sun, and X. Chu TEXTS-Diff: TEXTS-aware diffusion model for real-world text image super-resolution. In IEEE International Conference on Acoustics, Speech and Signal Processing, p. 4306–4310. Cited by: Table S1, Related Work. Heusel et al. (2017) M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems, Vol. 30, p. 6626–6637. Cited by: Evaluation Protocol and Metrics.. Ho et al. (2020) J. Ho, A. Jain, and P. Abbeel Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, Vol. 33, p. 6840–6851. Cited by: Introduction. Hu et al. (2022) E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen LoRA: low-rank adaptation of large language models. In International Conference on Learning Representations, Cited by: Related Work. Hu et al. (2025) Q. Hu, L. Fan, Y. Luo, Y. Yu, X. Guo, and Q. Fan Text-aware real-world image super-resolution via diffusion model with joint segmentation decoders. In Advances in Neural Information Processing Systems, Vol. 38, p. 61522–61543. Cited by: Table S1, Related Work. Kawar et al. (2022) B. Kawar, M. Elad, S. Ermon, and J. Song Denoising diffusion restoration models. In Advances in Neural Information Processing Systems, Vol. 35, p. 23593–23606. Cited by: Introduction. Kim et al. (2026) A. Kim, M. Jang, C. Moon, Y. Shin, J. Jeong, and S. Park FiDeSR: high-fidelity and detail-preserving one-step diffusion super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 38270–38280. Cited by: Table S1, Related Work, Latent Residual Correction. Kong et al. (2024) Y. Kong, W. Ma, L. Jin, and Y. Xue GARDEN: generative prior guided network for scene text image super-resolution. In International Conference on Document Analysis and Recognition, Vol. 14808, p. 196–214. Cited by: Related Work. Li et al. (2023) X. Li, W. Zuo, and C. C. Loy Learning generative structure prior for blind text image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 10103–10113. Cited by: Introduction, Related Work. Liang et al. (2021) J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, and R. Timofte SwinIR: image restoration using Swin transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, p. 1833–1844. Cited by: Related Work. Liu et al. (2025) B. Liu, Z. Yang, C. Chiu, and Y. Xiong TextDiff: enhancing scene text image super-resolution with mask-guided residual diffusion models. Pattern Recognition 164, p. 111513. Cited by: Related Work. Ma et al. (2023a) J. Ma, S. Guo, and L. Zhang Text prior guided scene text image super-resolution. IEEE Transactions on Image Processing 32, p. 1341–1353. Cited by: Introduction, Related Work. Ma et al. (2023b) J. Ma, Z. Liang, W. Xiang, X. Yang, and L. Zhang A benchmark for chinese-english scene text image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 19452–19461. Cited by: §D.3, Datasets.. Ma et al. (2022) J. Ma, Z. Liang, and L. Zhang A text attention network for spatial deformation robust scene text image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 5911–5920. Cited by: Introduction, Related Work. Min et al. (2026) J. Min, J. H. Kim, P. H. Cho, J. Lee, J. Park, M. Park, S. Kim, H. Park, and S. Kim Text-aware image restoration with diffusion models. In International Conference on Learning Representations, Cited by: Figure S2, Related Work. Niu et al. (2026) A. Niu, K. Zhang, Q. Yan, H. Jin, J. Sun, and Y. Zhang DualTSR: unified dual-diffusion transformer for scene text image super-resolution. External Links: 2603.14207 Cited by: Table S1, Related Work. Noguchi et al. (2024) C. Noguchi, S. Fukuda, and M. Yamanaka Scene text image super-resolution based on text-conditional diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, p. 1485–1495. Cited by: Related Work. Pan et al. (2025) C. Pan, X. Xu, G. Ding, Y. Zhang, W. Li, J. Xu, and Q. Wu Boosting diffusion-based text image super-resolution model towards generalized real-world scenarios. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, p. 7459–7468. Cited by: Table S1, Related Work, Confidence-Weighted Text Conditioning.. Park and Ko (2025) D. Park and S. P. Ko NCAP: scene text image super-resolution with Non-CAtegorical prior. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, p. 2432–2441. Cited by: Related Work. Rombach et al. (2022) R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 10684–10695. Cited by: Introduction. Saharia et al. (2023) C. Saharia, J. Ho, W. Chan, T. Salimans, D. J. Fleet, and M. Norouzi Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence 45 (4), p. 4713–4726. Cited by: Introduction. Singh et al. (2024) S. Singh, P. Keserwani, M. Iwamura, and P. P. Roy DCDM: diffusion-conditioned-diffusion model for scene text image super-resolution. In European Conference on Computer Vision, p. 303–320. Cited by: Related Work. Song et al. (2021) J. Song, C. Meng, and S. Ermon Denoising diffusion implicit models. In International Conference on Learning Representations, Cited by: Related Work. Tai et al. (2026) Y. Tai, R. Xie, C. Zhao, K. Zhang, Z. Zhang, J. Zhou, and J. Yang AddSR: accelerating diffusion-based blind super-resolution with adversarial diffusion distillation. Pattern Recognition 175, p. 113012. Cited by: Related Work. TomyEnrique et al. (2024) L. TomyEnrique, X. Du, K. Liu, H. Yuan, Z. Zhou, and C. Jin Efficient scene text image super-resolution with semantic guidance. In IEEE International Conference on Acoustics, Speech and Signal Processing, p. 3160–3164. Cited by: Related Work. Wang et al. (2024a) J. Wang, Z. Yue, S. Zhou, K. C. K. Chan, and C. C. Loy Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision 132 (12), p. 5929–5949. Cited by: Introduction. Wang et al. (2020) W. Wang, E. Xie, X. Liu, W. Wang, D. Liang, C. Shen, and X. Bai Scene text image super-resolution in the wild. In European Conference on Computer Vision, p. 650–666. Cited by: Introduction, Related Work. Wang et al. (2021) X. Wang, L. Xie, C. Dong, and Y. Shan Real-ESRGAN: training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, p. 1905–1914. Cited by: Appendix C, Datasets.. Wang et al. (2018) X. Wang, K. Yu, S. Wu, J. Gu, Y. Liu, C. Dong, Y. Qiao, and C. C. Loy ESRGAN: enhanced super-resolution generative adversarial networks. In European Conference on Computer Vision Workshops, p. 63–79. Cited by: Related Work. Wang et al. (2024b) Y. Wang, W. Yang, X. Chen, Y. Wang, L. Guo, L. Chau, Z. Liu, Y. Qiao, A. C. Kot, and B. Wen SinSR: diffusion-based image super-resolution in a single step. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 25796–25805. Cited by: Related Work. Wang et al. (2004) Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13 (4), p. 600–612. Cited by: Evaluation Protocol and Metrics.. Wei et al. (2025) B. Wei, Y. Zhou, L. Gao, and Z. Tang GlyphSR: a simple glyph-aware framework for scene text image super-resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39, p. 8277–8285. Cited by: Related Work. Wu et al. (2024) R. Wu, L. Sun, Z. Ma, and L. Zhang One-step effective diffusion network for real-world image super-resolution. In Advances in Neural Information Processing Systems, Vol. 37, p. 92529–92553. Cited by: Table S1, Related Work, Fixed-Timestep q-Sampling.. Xu et al. (2026) Z. Xu, X. Liu, Z. Chen, Y. Zhang, and X. Yang PRISM: prior rectification and uncertainty-aware structure modeling for diffusion-based text image super-resolution. External Links: 2605.13027 Cited by: Table S1, §E.2, Figure S2, Introduction, Related Work, Table 1. Yang et al. (2024) W. Yang, Y. Luo, M. Ibrayim, and A. Hamdulla More and less: enhancing abundance and refining redundancy for text-prior-guided scene text image super-resolution. In International Conference on Document Analysis and Recognition, Vol. 14808, p. 129–146. Cited by: Introduction. Ye et al. (2025) K. Ye, I. G. Dorado, M. Raptis, M. Delbracio, I. Zhu, P. Milanfar, and H. Talebi TextSR: diffusion super-resolution with multilingual OCR guidance. External Links: 2505.23119 Cited by: Related Work. Yuan et al. (2025) S. Yuan, R. Wang, K. Hao, X. Ma, C. Gao, L. Liu, and N. Sang StyleSRN: scene text image super-resolution with text style embedding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 18693–18702. Cited by: Related Work. Yue et al. (2023) Z. Yue, J. Wang, and C. C. Loy ResShift: efficient diffusion model for image super-resolution by residual shifting. In Advances in Neural Information Processing Systems, Vol. 36, p. 13294–13307. Cited by: Related Work. Zhang et al. (2021) K. Zhang, J. Liang, L. Van Gool, and R. Timofte Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 4791–4800. Cited by: Appendix C, Datasets.. Zhang et al. (2018) R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 586–595. Cited by: Evaluation Protocol and Metrics.. Zhang et al. (2024) Y. Zhang, J. Zhang, H. Li, Z. Wang, L. Hou, D. Zou, and L. Bian Diffusion-based blind text image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 25827–25836. Cited by: Table S1, Introduction, Related Work, Datasets.. Zhao et al. (2021) C. Zhao, S. Feng, B. N. Zhao, Z. Ding, J. Wu, F. Shen, and H. T. Shen Scene text image super-resolution via parallelly contextual attention network. In ACM International Conference on Multimedia, p. 2908–2917. Cited by: Introduction. Zhao et al. (2022) M. Zhao, M. Wang, F. Bai, B. Li, J. Wang, and S. Zhou C3-STISR: scene text image super-resolution with triple clues. In International Joint Conference on Artificial Intelligence, p. 1707–1713. Cited by: Related Work. Zhou et al. (2024) Y. Zhou, L. Gao, Z. Tang, and B. Wei Recognition-guided diffusion model for scene text image super-resolution. In IEEE International Conference on Acoustics, Speech and Signal Processing, p. 2940–2944. Cited by: Related Work. Zhu et al. (2023a) S. Zhu, Z. Zhao, P. Fang, and H. Xue Improving scene text image super-resolution via dual prior modulation network. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 37, p. 3843–3851. Cited by: Related Work. Zhu et al. (2023b) X. Zhu, K. Guo, H. Fang, R. Ding, Z. Wu, and G. Schaefer Gradient-based graph attention for scene text image super-resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 37, p. 3861–3869. Cited by: Related Work. Appendices This supplement provides implementation and evaluation details complementary to the self-contained main paper. It clarifies the system boundary relative to nearby diffusion methods, enumerates the modules executed by the one-step route, specifies the saved adaptation state, datasets, metrics, and controlled variants, and reports the exact values underlying the extended diagnostics. It also presents additional ablations, inference controls, uncertainty analysis, and qualitative comparisons. Appendix A System and Comparison Boundaries A.1 One-Step Inference Route For an input image, the VAE encoder first produces the unscaled three-channel latent zLRz_LR. TOLA uses the zero-indexed timestep t=999t=999 of the 1000-step linear schedule inherited from DiffTSR, whose endpoints are β0=0.0015 _0=0.0015 and β999=0.0205 _999=0.0205. One Gaussian sample is drawn to construct zt=α¯tzLR+1−α¯tϵ.z_t= α_tz_LR+ 1- α_tε. TransOCR is evaluated once. After Simplified-Chinese normalization, its token embeddings are multiplied by the aligned raw token confidences and passed to the frozen MoM together with 0.18215[zLR,zt]0.18215[z_LR,z_t]. MoM computes both inherited image and text conditions once, but the one-step IDM uses only the text condition as cross-attention context. IDM receives [zt,zLR][z_t,z_LR] and predicts noise once. Because the official DiffTSR checkpoint uses epsilon prediction, the initial clean latent is reconstructed as z^0=zt−1−α¯tϵ^α¯t. z_0= z_t- 1- α_t ε α_t. There is no reverse chain after this conversion. Latent residual correction (LRC) uses a compact correction branch to correct the LR-to-clean residual r=zLR−z^0r=z_LR- z_0: Δr=gϕ([zLR,r]),z^0,corr=zLR−(r+Δr). r=g_φ([z_LR,r]), z_0,corr=z_LR-(r+ r). Finally, the frozen decoder maps z^0,corr/0.18215 z_0,corr/0.18215 to xSRx_SR. Each image is restored once, without multi-sample selection or averaging. Appendix B Architecture and Saved State B.1 LoRA Placement and Frozen Components Rank-4 LoRA adapters use α=4α=4 and zero dropout. They are inserted into eligible linear and convolutional layers of IDM and the VAE encoder, including its encoder-side 1×11× 1 latent projection. The first input convolution of IDM, the decoder-side latent projection, and the complete VAE decoder remain unadapted. TransOCR, MoM, all base IDM/VAE weights, and the loaded but unused TDM decoder remain frozen. The conditioning TransOCR and the OCR-loss recognizer are separate frozen instances initialized from the same checkpoint. Trainable partition Parameters Saved? IDM LoRA 2,027,520 yes VAE-encoder LoRA 250,548 yes LRC 180,323 yes Total adaptation state 2,458,391 yes Table S3: Optimized and saved parameters. B.2 Latent Residual Correction TOLA implements LRC with a compact correction branch inspired by FiDeSR. It receives the six-channel concatenation [zLR,r][z_LR,r]. A 1×11×1 input projection maps 6 channels to 32, followed by one residual-in-residual dense block (RRDB) and a 1×11×1 output projection from 32 channels to 3. The RRDB contains three residual dense blocks; each residual dense block contains five densely connected 3×33×3 convolutions with growth width 16. Residual scaling is 0.2 at both dense-block and RRDB levels. The output projection is zero initialized, so LRC initially leaves z^0 z_0 unchanged. Small, Default, and Large LRC use (h,g,nRRDB)=(16,8,1)(h,g,n_RRDB)=(16,8,1), (32,16,1)(32,16,1), and (32,16,2)(32,16,2), respectively, with 0.045235M, 0.180323M, and 0.360323M parameters. The parameter-matched baseline uses a plain correction branch with a 6→306→30 projection, 11 two-convolution residual blocks with LeakyReLU and residual scaling 0.2, a 30→3030→30 mixing layer, and a zero-initialized 30→330→3 projection (0.180093M parameters). Appendix C Training Protocol Computing Environment. Experiments run on Ubuntu 22.04 with an AMD EPYC 9J14 96-Core Processor, 503 GiB of system memory, and NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs with 97,887 MiB per GPU. The NVIDIA driver version is 580.142. The primary environment uses Python 3.10.12, PyTorch 2.11.0+cu128, Torchvision 0.26.0+cu128, CUDA 12.8, cuDNN 9.19, PyIQA 0.1.15.post2, and LPIPS 0.1.4. The independent OCR evaluation uses PaddlePaddle 2.6.2 and PaddleOCR 2.7.3. These package versions define the software environment used for the reported results. Training reads 63,644 high-quality CTR preprocessing images and applies a deterministic local 98%/2% partition: 62,371 images for training and 1,273 for validation. This is a local partition rather than an official CTR split. Each image is degraded online with equal probability by the Real-ESRGAN or BSRGAN pipeline (Wang et al., 2021; Zhang et al., 2021); the scale is drawn from 1,2,4\1,2,4\. Both degraded inputs and targets use a 128×512128×512 canvas. The complete objective is ℒ=ℒ1(xSR,xHR)+ℒLPIPS-VGG(xSR,xHR)+0.02ℒOCR.L=L_1(x_SR,x_HR)+L_LPIPS -VGG(x_SR,x_HR)+0.02\,L_OCR. The frozen OCR recognizer remains differentiable with respect to its image input, so OCR-loss gradients reach LoRA and LRC. No latent ℓ1 _1, noise-prediction, score-distillation, VSD/CSD, or teacher-regularization loss is used. Method Image route Text route Distinction DiffTSR Iterative IDM TDM–MoM iterations Coupled image/text diffusion Boosting DiffTSR Iterative diffusion Confidence-guided TDM–MoM SR prior + progressive sampling OSEDiff One-step latent Degradation prompt LoRA + VSD FiDeSR One-step residual No OCR LRC + detail/frequency modules PRISM Single-pass flow Rectified prior Flow matching + structure modeling TADiSR Text-aware diffusion Text segmentation Joint segmentation decoders TEXTS-Diff One-step latent Abstract + region cues Real-Texts + text-aware restoration DualTSR Conditional flow Discrete diffusion Shared multimodal transformer TOLA (Ours) One IDM call TransOCR–MoM once One-Step; no TDM Table S1: Mechanistic positioning of the closest diffusion routes. Methods follow (Zhang et al., 2024; Pan et al., 2025; Wu et al., 2024; Kim et al., 2026; Xu et al., 2026; Hu et al., 2025; He et al., 2026; Niu et al., 2026). The table distinguishes system boundaries. Component Function Trainable update Calls/image TransOCR token and confidence prediction frozen 1 VAE encoder LR latent encoding LoRA 1 MoM text-aware condition construction frozen 1 IDM U-Net one-step diffusion prediction LoRA 1 LRC latent residual correction full branch 1 VAE decoder corrected-latent decoding frozen 1 Table S2: Inference boundary of TOLA. TransOCR, MoM, and IDM are each executed once; TDM is not executed. Item Setting Initialization official DiffTSR CTR checkpoint Training steps 100,000 GPUs 4× RTX PRO 6000 Batch/GPU 16 Effective batch 64 Precision FP32 Training time approximately 55 hours Optimizer AdamW Learning rate 5×10−55× 10^-5 Adam betas (0.9,0.999)(0.9,0.999) Weight decay 0.01 Warm-up 500 linear steps Gradient clipping 1.0 Selection validation performance Table S4: Primary training configuration. C.1 Hyperparameter Search and Selection Table S5 summarizes the search. The final architecture, objective, OCR-loss weight, and training timestep are selected using validation performance. Most candidate checkpoints use validation LPIPS; the no-LPIPS objective uses validation PSNR. The remaining optimization settings are reported in Table S4. Appendix D Dataset Appendix and Evaluation Metrics D.1 Dataset Inventory Table S6 summarizes the evaluation sets. The primary benchmark is the complete 8,089-image CTR-TSR-Test ×4× 4 set. All methods under the unified protocol are filename-matched to the same references and evaluated on a 128×512128×512 output canvas. D.2 Motivation and Coverage Existing benchmarks do not individually cover all conditions required for evaluating real scene-text restoration. CTR-TSR-Test provides a large paired benchmark but primarily reflects synthetic degradations, whereas available real-image datasets differ in language coverage, capture conditions, annotation format, and reference availability. RealCE-200 provides a paired protocol constructed from real multi-focal observations, while RT50 balances Chinese, English, and numeric text collected through heterogeneous acquisition routes. Together, these sets broaden the evaluated conditions. D.3 RealCE-200 Construction We construct RealCE-200 from the multi-focal images and official text annotations in the RealCE validation split before evaluation. For each official text annotation, identical bounding-box coordinates are applied to the corresponding 13m13\,m and 52m52\,m focal images, which serve as LR and reference observations. Both crops are bicubically resized to 128×512128×512 without geometric registration. Automatic filtering considers transcription length and composition, bounding-box size, aspect ratio, sharpness, and contrast. Duplicate candidates are removed using transcription identity and perceptual dHash. Official transcriptions are retained after whitespace cleaning, and no manual relabeling is performed. The resulting records use fixed identifiers R001–R200. RealCE is not used for training, validation, or checkpoint selection. Hyperparameter Values examined Count Final setting LoRA (r,α)(r,α) (2,2)(2,2), (4,4)(4,4), (8,8)(8,8) 3 (4,4)(4,4) LRC (h,g,nRRDB)(h,g,n_RRDB) (16,8,1)(16,8,1), (32,16,1)(32,16,1), (32,16,2)(32,16,2) 3 (32,16,1)(32,16,1) Reconstruction objective Full, without ℓ1 _1, without LPIPS 3 Full OCR-loss weight 0,0.02,0.05\0,0.02,0.05\ 3 0.02 Training timestep 899,949,999\899,949,999\ 3 999 Table S5: Hyperparameter and objective settings examined during development and the final configuration used by TOLA. Set Source Size Content and reference CTR-TSR-Test ×4× 4 Official CTR-TSR-Test 8,089 pairs Filename-aligned LR/HR images and transcriptions RealCE-200 RealCE multi-focal images (Ma et al., 2023b) 200 pairs Paired 13m13\,m LR/52m52\,m reference crops RT50 Multi-source real images 50 inputs 17 Chinese, 17 English, and 16 numeric samples Table S6: Evaluation datasets. RealCE-200 and RT50 are constructed in this work. D.4 RT50 Construction RT50 contains 50 real text inputs collected through multiple acquisition routes. The fixed composition contains 17 Chinese samples, 17 English samples, and 16 numeric samples; the difficulty partition contains 10 mild, 20 medium, and 20 hard samples. Candidate labels are restricted to 2–8 Chinese characters, 3–12 English letters, or 2–6 digits. Images must be at least 24×824×8 pixels with aspect ratios between 2 and 10. After canonical resizing, admissible candidates have gray mean in [20,235][20,235], gray standard deviation in [10,80][10,80], Laplacian variance in [2,1200][2,1200], edge density in [0.01,0.65][0.01,0.65], and entropy of at least 3.0. Repeated transcriptions are removed, and near duplicates are rejected at a dHash Hamming-distance threshold of 7. The selection protocol is finalized before model inference without consulting restoration or OCR outputs. Records use fixed identifiers RT001–RT050 and are normalized to the same 128×512128×512 model-input canvas. D.5 Included Dataset Records The public data release will include the exact image files evaluated in this work rather than requiring users to reconstruct them. For RealCE-200, it will provide all 200 LR/reference pairs together with the pair manifest, transcriptions, source filenames, and SHA-256 values. For RT50, it will provide all 50 input images together with the complete manifest, labels, provenance table, selection protocol, per-image geometry and quality statistics, category and difficulty assignment, dHash, and SHA-256 values. The accompanying records define the evaluated samples independently of model outputs and verify every released file. Upon publication, these exact datasets and their metadata will be publicly available under research-use terms documented with the release. D.6 Metric Implementations Predictions and references are matched by exact filename. PyIQA computes PSNR and SSIM on luminance without border crop. Evaluation LPIPS uses the AlexNet backbone, while training LPIPS uses VGG. DISTS and FID use PyIQA; FID is computed once over the complete output and reference directories. The primary recognizer is TransOCR. Predictions undergo full-width to half-width conversion, Simplified-Chinese conversion, and whitespace removal; case is preserved. Exact-match accuracy is denoted ACC. Per-image normalized edit-distance similarity is NED=1−ED(p,g)max(|p|,|g|,1),NED=1- ED(p,g) (|p|,|g|,1), and is averaged across the set. The independent cross-check uses PaddleOCR 2.7.3 with PaddlePaddle 2.6.2 and the Chinese P-OCRv4 recognition model; detection and orientation classification are disabled. This recognizer is not used for conditioning, training, validation, or checkpoint selection. Appendix E Extended Qualitative Comparisons E.1 Additional RealCE Comparison Figure S1 provides additional paired RealCE-200 examples under the unified evaluation protocol. E.2 BTL Cross-Protocol Comparison Figure S2 compares TOLA, DiffTSR, TeReDiff, and PRISM on five BTL-test ×4× 4 examples (Xu et al., 2026). The examples cover Chinese text, English words, and numeric sequences. Appendix F Controlled Variant Definitions All filename-paired comparisons use identical per-image Gaussian noise for the two compared outputs. Effect-size intervals use 5,000 filename-paired bootstrap resamples. Exact-match ACC is additionally tested with the exact paired McNemar test. Table S7 defines all architecture and conditioning controls. Figure S1: Qualitative comparison on five paired RealCE-200 examples. Rows show the LR input, SRCNN, ESRGAN, NAFNet, TSRN, TBSRN, TATT, MARCONet, DiffTSR, TeReDiff, Ours, and the HR reference. Figure S2: Cross-protocol comparison on five BTL-test ×4× 4 examples. Rows show the LR input, DiffTSR, TeReDiff (Min et al., 2026), PRISM (Xu et al., 2026), Ours, and GT. BTL-test combines real and rendered HQ sources with BSRGAN/Real-ESRGAN degradations; the PRISM row is reproduced from the original paper. Variant Definition Frozen one-step Frozen DiffTSR pathway with one IDM call and no LoRA or LRC. LRC only Default LRC is optimized while the pretrained image pathway remains frozen. LoRA only Rank-4 IDM and VAE-encoder LoRA without LRC. Full Rank-4 LoRA plus the default LRC and confidence-aware MoM condition. No MoM The TransOCR–MoM conditioning branch is removed during training and inference. Direct tokens Predicted token embeddings are supplied without MoM latent fusion. Null context Fixed-checkpoint control retaining MoM with a null text context. Uniform confidence Fixed-checkpoint control retaining predicted tokens and MoM with unit confidences. GT prior Fixed-checkpoint oracle replacing predicted tokens with ground-truth text. IDM LoRA only VAE-encoder LoRA is omitted; IDM LoRA and default LRC remain. VAE LoRA only IDM LoRA is omitted; VAE-encoder LoRA and default LRC remain. VAE LoRA off Inference disables all VAE-encoder LoRA, including the latent projection, while retaining IDM LoRA and LRC. Table S7: Architecture and conditioning controls used in the diagnostics. Configuration PSNR SSIM LPIPS DISTS FID Full 24.8095 .7235 .2288 .1953 12.6849 Rank 2 23.9287 .6979 .2835 .2168 16.8514 Rank 8 23.9573 .6993 .2815 .2179 17.2784 LRC Small 23.9506 .6984 .2870 .2192 17.4977 LRC Large 23.9967 .6990 .2781 .2154 15.6799 No ℓ1 _1 23.1142 .6906 .2801 .2153 15.6182 No LPIPS 24.0026 .6888 .3820 .2744 50.5338 OCR wt. 0 23.9157 .6997 .2741 .2132 16.0079 OCR wt. .05 23.9290 .6961 .2955 .2232 18.1809 Train t=899t=899 24.2532 .7088 .2571 .2101 13.5299 Train t=949t=949 24.1818 .7068 .2644 .2121 13.8379 Table S8: Image-quality diagnostics on CTR-TSR-Test ×4× 4. The reference uses rank 4, the default 0.180M LRC, OCR weight 0.02, and t=999t=999. Configuration T-ACC T-NED P-ACC P-NED Full .5401 .7370 .4626 .6585 Rank 2 .5138 .7056 .4253 .6247 Rank 8 .5119 .7052 .4211 .6224 LRC Small .5079 .7038 .4235 .6235 LRC Large .5097 .7054 .4247 .6240 No ℓ1 _1 .5092 .7031 .4248 .6237 No LPIPS .5044 .6987 .4015 .6049 OCR wt. 0 .5025 .6981 .4253 .6216 OCR wt. .05 .5096 .7042 .4217 .6234 Train t=899t=899 .5127 .7063 .4249 .6266 Train t=949t=949 .5107 .7060 .4271 .6256 Table S9: Recognition diagnostics for the same configurations. Most checkpoints use validation LPIPS; the no-LPIPS variant uses validation PSNR. Inference Runs. The main comparisons, ablations, and controls use one restored output per image without multi-sample selection or output averaging. The stochastic-stability analysis contains three inference runs and reports their mean and standard deviation. Bootstrap resampling does not generate additional restored outputs. Appendix G Extended Ablation and Control Results G.1 Objective and Hyperparameter Diagnostics Tables S8 and S9 show that rank 4 provides a better overall balance than ranks 2 and 8. The default LRC also outperforms its smaller and larger variants overall, indicating that additional capacity is not the source of the gain. Removing ℓ1 _1 causes the largest PSNR reduction, whereas removing LPIPS most strongly degrades perceptual and distributional metrics. An OCR weight of 0.02 outperforms both zero weight and 0.05, and t=999t=999 gives the strongest overall training result among the tested timesteps. G.2 Prior-Stratified Performance Table S10 separates samples according to whether the predicted TransOCR prior matches the ground truth. Statistic Correct prior Wrong prior N 3,439 4,650 Mean confidence 0.9536 0.6015 DiffTSR ACC 0.9514 0.1043 TOLA ACC 0.9683 0.2217 DiffTSR NED 0.9872 0.4038 TOLA NED 0.9906 0.5482 TOLA PSNR 26.2307 23.7535 TOLA LPIPS 0.1652 0.2760 Table S10: Performance stratified by TransOCR prior correctness, with image and recognition metrics. G.3 Inference Controls and Stochastic Stability Table S11 shows that t=949t=949 slightly increases PSNR and SSIM but worsens LPIPS, DISTS, and FID, while t=899t=899 further degrades perceptual and distributional quality. The default t=999t=999 therefore provides the best overall balance. Zero noise and disabled VAE LoRA substantially reduce reconstruction quality, and the input controls rule out copying or a frozen-VAE round trip as explanations for the gains. Repeated inference shows low sensitivity to the sampled Gaussian noise. Efficiency Measurement Details. All methods are timed on the same 100 CTR-TSR-Test images at batch size 1 on one NVIDIA RTX PRO 6000 GPU. Timing follows warm-up, excludes loading, and synchronizes CUDA. TOLA timing includes VAE encoding, TransOCR, MoM, one IDM evaluation, LRC, and VAE decoding; DiffTSR and TeReDiff use their full 200- and 50-step routes. MACs sum restoration-backbone evaluations at 128×512128×512. Counts exclude VAEs and auxiliary recognition, detection, and spotting modules, so the reported 876.23M restoration-module count differs from the 2.46M optimized and saved adaptation state. Planned Public Release. Upon publication, we will release the implementation, configurations, and exact RealCE-200 and RT50 image records under research-use terms. The release will document the required third-party checkpoints without redistributing their pretrained weights. G.4 Paired Uncertainty Table S12 reports paired intervals, which exclude zero for the LRC effects on PSNR, SSIM, LPIPS, and DISTS, but not NED; the exact paired McNemar test likewise finds no significant ACC difference. Because all comparisons use filename-aligned outputs, the intervals measure within-sample changes and separate consistent reconstruction gains from limited recognition changes. Paired resampling preserves image correspondence when estimating uncertainty across methods. Predicted conditioning significantly improves ACC over null context, whereas the GT-prior gap confirms remaining headroom from prior errors. Figure S3 plots these conditioning controls, and Figure S4 visualizes the paired LRC effects. (a) Fixed-checkpoint image metrics Control PSNR↑ SSIM↑ LPIPS↓ DISTS↓ FID↓ Inference t=999t=999 24.8067 .7235 .2289 .1953 12.6849 Inference t=949t=949 24.9307 .7257 .2536 .2109 17.8904 Inference t=899t=899 24.6970 .7219 .2777 .2250 22.6079 ϵ= = 0 23.4153 .7014 .2360 .2011 14.7950 VAE LoRA off 23.0239 .7188 .2458 .2074 17.4872 (b) Fixed-checkpoint recognition metrics Control T-ACC↑ T-NED↑ P-ACC↑ P-NED↑ Inference t=999t=999 .5391 .7362 .4626 .6585 Inference t=949t=949 .5402 .7381 .4614 .6589 Inference t=899t=899 .5395 .7365 .4615 .6585 ϵ= = 0 .5359 .7295 .4539 .6521 VAE LoRA off .5389 .7335 .4572 .6552 (c) Input controls Control PSNR SSIM LPIPS DISTS T-ACC T-NED LR input 21.9863 .5799 .6519 .3964 .4235 .6102 Frozen VAE 15.5357 .5033 .5913 .4025 .3427 .5245 (d) Repeated stochastic inference Metric LoRA LoRA+LRC PSNR 24.4242±.001924.4242±.0019 24.8082±.001324.8082±.0013 SSIM .710845±.000034.710845±.000034 .723514±.000012.723514±.000012 LPIPS .241043±.000078.241043±.000078 .228854±.000054.228854±.000054 DISTS .197032±.000057.197032±.000057 .195314±.000031.195314±.000031 T-ACC .537603±.000143.537603±.000143 .540487±.001192.540487±.001192 T-NED .734259±.000461.734259±.000461 .736198±.000637.736198±.000637 Table S11: Inference controls and stochastic stability on CTR-TSR-Test ×4× 4. Panels (a)–(c) use fixed checkpoints; panel (d) reports mean± deviation across repeated inference. VAE LoRA off disables all VAE-encoder LoRA, including the latent projection, while retaining IDM LoRA and LRC. Comparison Metric Change 95% CI LRC–LoRA PSNR +0.3847+0.3847 [+0.3740,+0.3955][+0.3740,+0.3955] SSIM +0.0127+0.0127 [+0.0123,+0.0130][+0.0123,+0.0130] LPIPS −0.0121-0.0121 [−0.0128,−0.0115][-0.0128,-0.0115] DISTS −0.0017-0.0017 [−0.0021,−0.0013][-0.0021,-0.0013] ACC (p=0.584p=0.584) +0.0014+0.0014 [−0.0031,+0.0058][-0.0031,+0.0058] NED +0.0025+0.0025 [−0.0002,+0.0053][-0.0002,+0.0053] Pred.–Null ACC (p=0.0035p=0.0035) +0.0083+0.0083 [+0.0028,+0.0138][+0.0028,+0.0138] NED +0.0082+0.0082 [+0.0051,+0.0113][+0.0051,+0.0113] GT–Pred. ACC (p<0.001p<0.001) +0.0862+0.0862 [+0.0797,+0.0925][+0.0797,+0.0925] NED +0.0716+0.0716 [+0.0675,+0.0757][+0.0675,+0.0757] Table S12: Filename-paired uncertainty on CTR-TSR-Test ×4× 4. Effect-size intervals use 5,000 paired bootstrap resamples; ACC significance is additionally evaluated with the exact paired McNemar test. Changes are computed from unrounded values. G.5 Exact Recognition Diagnostics Tables S13–S16 summarize recognition behavior by confidence, text category, and edit operation. Confidence denotes the mean confidence of the predicted TransOCR tokens. Prior ACC is the exact-match accuracy of the predicted text prior, and D and T denote DiffTSR and TOLA, respectively. The same 8,089 filename-aligned CTR-TSR-Test ×4× 4 samples and TransOCR normalization are used throughout. The confidence relationship is monotonic, but the method improvement is not restricted to high-confidence priors. TOLA improves ACC and NED in all four confidence intervals and all four text categories. The edit decomposition also shows reductions in substitutions, deletions, and insertions, lowering the overall error rate from 42.34 to 32.55 edits per 100 characters. The large gap between correct- and wrong-prior strata identifies prior recognition as the primary remaining source of exact-transcription error. Confidence N Prior D-A T-A D-N T-N [0,.25)[0,.25) 246 .0041 .0407 .1382 .1618 .3454 [.25,.50)[.25,.50) 1,344 .0074 .0476 .1339 .1976 .3694 [.50,.75)[.50,.75) 1,926 .0685 .1506 .2721 .4273 .5771 [.75,1][.75,1] 4,573 .7208 .7420 .7923 .9063 .9321 Table S13: Exact recognition by predicted-prior confidence. D/T denote DiffTSR/TOLA; A/N denote ACC/NED. Category N D-A T-A D-N T-N Chinese 6,946 .4587 .5346 .6366 .7251 English 884 .5204 .5792 .7539 .8098 Mixed 120 .1750 .2917 .5543 .6435 Num./Other 139 .6475 .7266 .8487 .9038 Table S14: Exact recognition by text category. TOLA improves over DiffTSR in every category. Method Sub. Del. Ins. Total Edits/100 DiffTSR 14,657 3,139 673 18,469 42.34 TOLA 12,082 1,529 589 14,200 32.55 Table S15: Character-level edits over 43,622 ground-truth characters under the normalized TransOCR evaluation used for ACC and NED. Prior N Exact GT Copy Other Correct 3,439 3,330 – 109 Wrong 4,650 1,031 376 3,243 Table S16: Behavior conditioned on predicted-prior correctness. Counts sum to all 8,089 CTR-TSR-Test samples. Figure S3: Text-prior controls with null (N), uniform (U), predicted (C), and ground-truth oracle (GT) conditions. Figure S4: Paired LRC effects over LoRA-only with 95% CIs.