acceptodds
Under review as a conference paper at ICLR 2027

Distill the Prior, Drop the Diffusion: Compact Feed-Forward Text Restoration from Diffusion Priors

Abstract

Diffusion models provide powerful generative priors for image restoration, but transferring such knowledge beyond the diffusion architecture remains challenging. Can a restoration prior learned by a large diffusion model be absorbed by a fundamentally different feed-forward network, despite the mismatch in internal representations and inference mechanisms? Moreover, under aggressive model compression, can this prior still be retained rather than collapsing with student capacity? We study these questions in text image super-resolution and introduce PixelSTR, a diffusion-prior distillation framework that transfers restoration behavior entirely in output space. Instead of aligning diffusion latents or intermediate features, a lightweight pixel-space student learns the appearance and local structure of teacher outputs through RGB and spatial-gradient supervision, while paired high-resolution images and transcriptions provide reconstruction and recognition constraints. PixelSTR achieves over 100 parameter compression relative to the teacher restoration module. Under matched training, diffusion-derived supervision improves PSNR by 1.51 dB and OCR accuracy by at least 6.70 percentage points over supervised-only training, while replacing the diffusion targets with additional ground-truth targets does not reproduce these gains. These results suggest that generative restoration priors can be retained without retaining the generative model.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
text image super-resolution, diffusion priors, knowledge distillation, efficient image restoration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_xehbOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PixelSTR, a lightweight distillation framework for text image super-resolution that aims to transfer image restoration priors from a large diffusion model to a non-diffusion feed-forward network, thereby reducing inference costs. Specifically, the method employs a frozen diffusion model as the teacher and trains a lightweight NAFNet student through RGB and spatial-gradient matching, together with reconstruction, perceptual, and OCR supervision based on ground-truth high-resolution images and text annotations. Experimental results demonstrate that PixelSTR achieves approximately a 15.8× forward-pass speedup while using only 0.98% of the teacher's restoration-module parameters, maintaining competitive image restoration quality and text recognition performance.

Critical strengths
  1. Comprehensive experiments demonstrate a favorable efficiency–performance trade-off. PixelSTR achieves a 15.8× forward-pass speedup using only 0.98% of the teacher's restoration-module parameters, while maintaining competitive image restoration quality and text recognition performance.

  2. Simple and practical framework design.

Critical weaknesses
  1. It remains unclear why completely abandoning the diffusion is a necessary or sufficiently well-motivated design choice, rather than further compressing and accelerating diffusion models themselves. The authors should provide a clearer justification for this fundamental design decision and explain the distinct advantages of their approach over retaining a compact diffusion-based architecture.

  2. Why do RGB teacher supervision and gradient teacher supervision, which appear to be two relatively independent loss terms, yield almost no improvement when applied individually, yet lead to substantial performance gains when combined? Although the paper demonstrates this phenomenon experimentally, it does not provide a sufficient explanation for the underlying mechanism.

  3. Limited novelty. The main contribution of this work lies in distilling image restoration priors from a large diffusion model into a lightweight, non-diffusion feed-forward network, thereby significantly reducing inference costs while maintaining competitive restoration performance. However, prior work DTKD [1] has already investigated cross-architecture Diffusion-to-Non-Diffusion distillation for image super-resolution, transferring generative priors from a diffusion model to an efficient Transformer student through frequency-domain knowledge distillation. Therefore, the authors should more clearly articulate the key technical innovations of their method compared with DTKD.

[1] DTKD: Diffusion-to-Transformer Heterogeneous Knowledge Distillation for Efficient and Perceptually Enhanced Super-Resolution

Questions for authors

Please see weakness section.

1 comment
AuthorsOpenReview ↗

Response to Reviewer xehb

W1. Why deploy a feed-forward student? At 8.62M total parameters, PixelSTR gains 3.8720 dB and 13.69/15.77 pp in TransOCR/PaddleOCR (T/P) over pruned TOLA on CTR (n=8,089). L2 channel pruning narrows the denoiser, VAE, OCR and MoM. With the intact latent refiner, the count is 8.624603M, 16 fewer parameters than PixelSTR. Pruned TOLA retains single-step diffusion and uses TOLA’s recipe for 50K recovery updates. PixelSTR uses GT and RGB/gradient KD for 50K steps and leads TinySR on all five metrics.

ModelParams MPSNR↑SSIM↑LPIPS↓T-ACC%↑P-ACC%↑
Pruned TOLA (diffusion), 50K recovery8.62460320.06530.56300.655335.4324.07
TinySR362.82622221.60930.53410.591642.8931.65
PixelSTR, 50K8.62461923.93730.68630.291449.1239.84

Efficiency. PixelSTR uses the fewest parameters and MACs and runs 15.8× faster than TOLA.

MethodParams (M)↓MACs (G)↓Reported time (ms/image)↓
TOLA876.23297.497131.90
OSEDiff1765.30726.18767.09
TinySR362.832356.823139.72
PixelSTR8.624.9528.36

Real images. PixelSTR leads bicubic, OSEDiff and TinySR on RealCE-1K (1,000 crops, MARCONet++ [1]) and gains 1.0019 dB over OSEDiff on CTR.

MethodPSNR↑SSIM↑T-ACC%↑P-ACC%↑
Bicubic19.70950.621142.9053.00
OSEDiff20.11970.634447.7062.30
TinySR20.10590.610747.9062.20
PixelSTR (ours)20.57860.669151.3066.40

W2. Why do RGB and gradient targets help together? Together they constrain stroke intensity and adjacent-pixel changes. Matching finite differences alone allows a constant offset. Matching both lowers edge, non-edge and image-gradient MAE by 10.48%, 31.71% and 16.60% versus GT alone. All models use the same HR edge mask per CTR image (top 25% by finite-difference magnitude).

SupervisionEdge MAE↓Non-edge MAE↓Image-gradient MAE↓
GT only0.1080270.0646310.030743
RGB only0.1074800.0637720.029618
Gradient only0.1079800.0645320.029919
RGB + gradient0.0967030.0441340.025639

Joint KD improves word recognition. It reads 630 strings both single-term students miss and corrects 749 GT-only errors (net gain 542). The joint gain exceeds the sum of the individual gains by 1.4299 dB and 6.57 pp in TransOCR accuracy.

The joint student has better-aligned KD gradients. Their mean cosine is 0.4484 for the joint student and 0.0137 for RGB only. Median norm ratios differ from 1 by 0.03%–1.53%. On eight fixed batches of 16, we sum the RGB and 0.1-weighted finite-difference parameter gradients and divide the sum’s L2 norm by the RGB-gradient norm.

Student stateMedian joint/RGB norm ratioMean KD cosine
Shared initialization1.0054570.1148
Supervised only1.0067440.1593
RGB only1.0002760.0137
Gradient only1.0043300.1291
RGB + gradient1.0153040.4484

GT and teacher update directions conflict. A negative gradient cosine means teacher descent increases GT loss to first order. Of 128 inputs, GT opposes RGB on 18 (14.06%), finite differences on 29 (22.66%) and joint KD on 18 (14.06%) (Appendix I.2, Fig. 10, p. 22). RGB-L1 derivatives oppose where (S−H)(S−T)<0(S-H)(S-T)<0 (S student, H reference, T teacher), averaging 34.69% of channel positions. In Appendix I.2, Fig. 11, p. 22, this fraction is 36.23% and GT/joint-KD cosine is −0.1649.

W3. What does PixelSTR add beyond DTKD? PixelSTR distills stroke intensity and edges by matching teacher RGB and adjacent-pixel differences at corresponding locations (fixed weights 1.0/0.1). Transcripts supervise character identity. DTKD [2] targets perceptual quality in natural images, emphasizes mid-frequency wavelet bands and increases the KD weight during training.

CTR recognition and restoration. Both recognizers rank PixelSTR above the two wavelet controls in the CTR table. Against the fixed-weight control, the T/P gains are 1.38 pp and 3.74 pp, respectively. PSNR rises 0.8973 dB and LPIPS falls 0.1060. Fixed beats linear by 0.0981 dB and 0.57 pp P-ACC.

ObjectivePSNR↑SSIM↑LPIPS↓T-ACC%↑P-ACC%↑
DTKD-style linear22.94190.61310.393647.5335.53
DTKD-style fixed23.04000.61710.397447.7436.10
PixelSTR23.93730.68630.291449.1239.84

Controlled DTKD-style wavelet baselines. We keep TOLA, the 8.62M NAFNet student, CTR data, GT losses (L1, LPIPS, OCR), optimizer, batch 16 and 50K updates. Two-level Haar KD replaces RGB/gradient KD. Low/mid/high weights are 0.25/0.50/0.25. The KD weight rises linearly from 0 to 1 or stays at 1.

References

  • [1] Li et al. MARCONet++. TPAMI 2025.
  • [2] Park and Song. DTKD. Electronics 2026.
Reviewer_if4LOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies whether the restoration behavior of a large diffusion model can be transferred to a small feed-forward network for text image super-resolution (×4). The teacher is TOLA, a frozen one-step diffusion model whose restoration module has about 876M parameters. During training it generates RGB targets online with fresh noise at every iteration. The student is a NAFNet with 8.62M parameters that operates on bicubic-upsampled 512×128 inputs. Information is transferred only in output space: an L1 loss on the teacher output plus a 0.1-weighted L1 loss on its finite-difference gradients. These are added to standard GT supervision (L1, LPIPS-VGG, and a 0.02-weighted cross-entropy loss from a frozen TransOCR recognizer).

On CTR-TSR-Test the student reaches 23.94 dB PSNR and 0.491 TransOCR accuracy (teacher: 24.81 / 0.540) and beats nine retrained baselines. It uses about 1% of the teacher's restoration parameters and runs about 15.8× faster. The main controlled evidence comes from two comparisons:

Full vs. supervised-only training (+1.51 dB PSNR, +6.70 pp TransOCR accuracy), with 3 seeds each.

A GT-target control and a TT/TG/GT/GG target-source factorial.

Zero-shot results on 200 RealCE crops are also reported.

Critical strengths
  1. Practical efficiency with competitive quality. On CTR, an 8.6M-parameter single-pass model comes close to its one-step diffusion teacher and outperforms the diffusion-based TextSR methods DiffTSR and TeReDiff on all reported metrics, at a small fraction of their cost. This is a useful engineering result for deployment.
  2. Careful evaluation practice. The paper uses an independent OCR evaluator (PaddleOCR) that is not involved in training, repeated runs for the main comparison, and paired bootstrap intervals. It also reports limitations honestly, including the bicubic baseline on RealCE. These choices make the headline Full vs. supervised-only gap credible.
Critical weaknesses
  1. The method lacks novelty: it is a straightforward output-level distillation that uses the teacher's restoration as a soft target. The only distillation-specific component is an L1 loss on the one-step diffusion teacher's RGB output, plus a 0.1-weighted L1 loss on its gradients. The remaining losses (L1, LPIPS, and OCR cross-entropy against the ground truth) are standard in text super-resolution. Training a compact network on the outputs of a stronger model is a well-established form of knowledge distillation. Moreover, nothing in the method is specific to a diffusion teacher. It does not exploit the teacher's stochasticity or confidence, and it does not filter incorrect targets, even though the teacher fails on about 46% of CTR samples by TransOCR. The practical gains are useful, but the technical contribution is too limited for ICLR.

  2. The comparison with prior methods is not convincing because the baseline training protocol is not reported and several results are anomalous. The paper states only that the non-diffusion baselines were retrained on the CTR split. It gives no details on their degradation pipeline, input resolution, losses, or training budget. Several results suggest the baselines were not trained under comparable conditions. Recognition-oriented methods (TSRN, TBSRN, TATT) reach only 0.33–0.36 TransOCR accuracy, below the near-bicubic SRCNN (0.43). The 67.9M NAFNet baseline (21.56 dB) is outperformed by the paper's own 8.6M NAFNet trained only with GT losses (22.43 dB), so part of the reported margin likely comes from the training configuration rather than from distillation. No bicubic reference is given on CTR. As a result, the claim of outperforming all nine baselines is difficult to assess.

  3. Comparisons with the most relevant methods are missing. The paper aims at diffusion-level quality with low inference cost, yet it does not compare against methods that pursue the same goal. These include one-step diffusion SR (e.g., OSEDiff, SinSR, AddSR), compact diffusion restorers (e.g., TinySR, PocketSR), and existing knowledge-distillation methods for SR. Several of these are cited in the related work but never evaluated. The current baselines (mostly earlier TextSR networks and two multi-step diffusion models) cannot show that output-level distillation is a better route than these alternatives.

Questions for authors
  1. Role of the diffusion teacher (W1). Is the gain specific to a diffusion teacher? A control using a strong non-diffusion restorer trained on CTR as the teacher, with the same loss, would clarify whether the method transfers a diffusion prior or performs generic output-level distillation.

  2. Baseline protocol (W2). Please report the full training protocol for each retrained baseline (degradation pipeline, input resolution, losses, training steps), add a bicubic reference on CTR-TSR-Test, and explain why recognition-oriented methods (TSRN, TBSRN, TATT) achieve lower recognition accuracy than SRCNN. Would the conclusions still hold if the baselines were trained with the same losses as the student (L1 + LPIPS + OCR)?

  3. Missing comparisons (W3). How does PixelSTR compare with one-step diffusion SR methods (e.g., OSEDiff, SinSR) and existing KD-for-SR methods under the same data, input protocol, and a comparable parameter or latency budget?

1 comment
AuthorsOpenReview ↗

Response to Reviewer if4L

W1/Q1. Distilling appearance and stroke structure. PixelSTR transfers a one-step diffusion teacher’s RGB values and local stroke transitions at matching pixels to an 8.62M feed-forward student. HR anchors appearance and the transcript guides character identity. Only the student runs at inference. Joint targets give 23.9373 dB and 49.12% TransOCR accuracy, versus 22.5012 dB/42.43% for RGB alone and 22.4352 dB/42.54% for gradients alone. A non-diffusion teacher also helps. Replacing capacity-matched NAFNet with TOLA adds 0.2004 dB.

The student corrects teacher errors. Table 10 groups images by TransOCR on cached teacher outputs. On the 3,708 it misreads, student accuracy is 4.75% with distillation and 3.24% with GT alone.

Teacher OCRImagesGT-only T-ACC%PixelSTR T-ACC%Gain, pp [95% CI]
Correct4,38175.5886.67+11.09 [9.91, 12.30]
Incorrect3,7083.244.75+1.51 [0.76, 2.27]

GT transcripts guide character recognition. The student uses the true transcript even when the teacher renders a wrong character. HR L1 and LPIPS guide appearance. Frozen OCR scores text (cross-entropy weight 0.02). We use fresh teacher noise and every output, weighting RGB and finite-difference L1 by 1 and 0.1. PixelSTR repeats 15.10% of nonempty wrong teacher strings versus 9.60% for GT alone.

W2/Q2. (1) Baseline comparison. We are very confident that our comparison is convincing. We retrained all nine original diffusion and non-diffusion baselines, including DiffTSR and TeReDiff, in the same training environment on CTR for 100K updates each. The split has 62,371 training and 1,273 held-out images. We sampled Real-ESRGAN or BSRGAN equally on 512×128 HR images at scales 1, 2 or 4. All used the same 8,089 CTR ×4 test pairs and metric code.

(2) Baseline results. The data are normal. On CTR, DiffTSR and Boosting DiffTSR [1] likewise report 60.31%/51.23% OCR for SRCNN/MARCONet and 20.74/19.41 dB for SRCNN/TSRN. PRISM [2, Table 1] reports 31.37%/33.41%/29.80% PP-OCRv5 for TSRN/TBSRN/TATT on BTL-test ×4 (CTR, SA-Text and synthetic text).

SRCNN versus bicubic. SRCNN exceeds bicubic by 10.21/10.39 percentage points (pp) in TransOCR/PaddleOCR accuracy (T/P). On CTR, TSRN/TBSRN/TATT score 33.07%/35.39%/35.58% TransOCR.

CTR inputTransOCR ACC%↑PaddleOCR ACC%↑
Bicubic32.3422.64
SRCNN42.5533.03

(3) Does the ranking hold with matched GT losses? Yes. PixelSTR leads on all four metrics. It gains 2.0192/2.6788 dB over NAFNet/TSRN. Its T/P gains are 7.58/8.38 pp over NAFNet and 16.10/17.01 pp over TSRN. Original NAFNet (67.89M) scored 21.5568 dB vs. 22.4290 for the GT-only student. Both baselines used L1 + LPIPS + 0.02 OCR for 50K updates (batch 16, FP32 AdamW at 2e-4, 500 warm-up steps). On the same 8.62M student, KD adds 1.5083 dB and 6.70/8.03 pp in T/P (Tables 4–5).

ModelPSNR↑LPIPS↓T-ACC%↑P-ACC%↑
NAFNet21.91810.563141.5431.46
TSRN21.25850.560533.0222.83
PixelSTR23.93730.291449.1239.84

W3/Q3. How does PixelSTR compare? PixelSTR uses fewer parameters than OSEDiff and TinySR. The tabulated rates favor it over both and the DTKD-style controls [3] on four CTR metrics, gaining 1.0019 dB over OSEDiff and 0.8973 dB over fixed DTKD, with T/P gains of 1.38 pp and 3.74 pp, respectively. These controlled DTKD-style wavelet baselines replace RGB/gradient KD with a two-level Haar loss. They share the TOLA teacher, 8.62M NAFNet student, GT losses (L1, LPIPS, OCR), optimizer, batch 16 and 50K steps.

MethodPSNR↑LPIPS↓T-ACC%↑P-ACC%↑
OSEDiff22.93540.362343.0332.50
TinySR21.60930.591642.8931.65
DTKD-style linear22.94190.393647.5335.53
DTKD-style fixed23.04000.397447.7436.10
PixelSTR23.93730.291449.1239.84

Inference efficiency. PixelSTR leads in parameters and MACs. Its 8.36 ms forward pass is 15.8× faster than TOLA’s.

MethodParams (M)↓MACs (G)↓Reported time (ms/image)↓
TOLA876.23297.497131.90
OSEDiff1765.30726.18767.09
TinySR362.832356.823139.72
PixelSTR8.624.9528.36

RealCE-1K. PixelSTR leads on all four metrics across 1,000 crops from MARCONet++ [4].

MethodPSNR↑SSIM↑T-ACC%↑P-ACC%↑
Bicubic19.70950.621142.9053.00
OSEDiff20.11970.634447.7062.30
TinySR20.10590.610747.9062.20
PixelSTR (ours)20.57860.669151.3066.40

References

Reviewer_CN9gOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PixelSTR, a compact feed-forward text restoration model trained to distill RGB outputs and spatial gradients from a frozen, one-step TOLA diffusion teacher. The student combines this supervision with paired reconstruction, perceptual, and OCR losses, then restores an image in a single forward pass. The approach is practical, and the reported results suggest that useful restoration behavior can be transferred to a much smaller model. My main reservation is that the technical contribution appears relatively straightforward, while the real-world and challenging-case evidence is not yet strong enough to establish the method’s broader value.

Critical strengths
  1. The task is practically relevant: the method aims to retain some benefits of a diffusion-based restorer while avoiding its inference cost.

  2. The student is lightweight (8.62M parameters), and the reported ablations indicate that teacher distillation improves over supervised-only training.

  3. The paper includes useful ablations and repeatability analysis, making the reported synthetic-benchmark results reasonably well supported.

Critical weaknesses
  1. The specific value of a diffusion teacher is not isolated. The central idea is to distill the teacher’s RGB outputs and spatial gradients into a feed-forward student. The results show that this supervision helps, but do not establish that the benefit comes specifically from a diffusion prior rather than from distilling a strong teacher in general. Please compare against a capacity- and data-matched non-diffusion teacher, using the same student, losses, and training budget.

  2. Real-world evidence is limited. RealCE contains only 200 crops from 73 source images, and PixelSTR’s gains over bicubic are modest. Please evaluate on a larger and more diverse real-world set, report source-image-aware statistics, and clarify how much the results vary across source images.

  3. CTR image-quality handling is unclear. The paper describes resizing and synthetic degradation, but does not clearly explain whether the underlying CTR images are screened for source quality or how quality variation is handled; this is also difficult to assess from Figure 3. Since low-quality or inconsistent source images may affect both the restoration targets and OCR evaluation, please specify the filtering and preprocessing protocol and report the data-quality distribution or results stratified by source quality.

  4. The qualitative examples do not establish the method’s operating boundary. The shown examples do not appear sufficiently challenging, and the teacher’s advantage over DiffTSR is not consistently evident in them. Please include harder, matched examples—such as severe blur, compression, or low recognizer confidence—and report performance stratified by degradation severity. Comparisons should include the low-resolution input, DiffTSR, the teacher, and PixelSTR, along with representative failure cases.

Questions for authors

See the weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer CN9g

W1. Diffusion prior. Replacing NAFNet with diffusion teacher TOLA adds 0.2004 dB to student PSNR and 1.04/0.71 pp to TransOCR/PaddleOCR (T/P) accuracy. Non-diffusion NAFNet gains 1.3079 dB over GT alone. TOLA beats StyleSRN [1] on all five metrics.

Student supervisionPSNR↑LPIPS↓FID↓T-ACC%↑P-ACC%↑
GT only22.42900.472286.294642.4231.81
StyleSRN teacher23.67300.316531.337047.6239.10
NAFNet teacher23.73690.306728.218248.0839.13
TOLA teacher (PixelSTR)23.93730.291423.859549.1239.84

Controls. TOLA/NAFNet have 876.23M/876.21M restoration parameters and 100K steps on the same CTR split. All use the same 8.62M student, 50K steps, batch 16, GT L1+LPIPS-VGG+0.02 OCR CE and RGB/gradient KD (1.0/0.1). Test n=8,089.

W2. RealCE-1K. PixelSTR improves all six metrics over bicubic. Across 1,000 crops from 192 photos (MARCONet++ [2]), zero-shot gains are 0.8691 dB in PSNR, 0.0480 in SSIM and 8.40/13.40 pp in T/P. RealCE-200 used 200 handpicked hard crops from 73 photos. The sets share 81 crops and 58 photos. The 134 new photos contribute 674 crops.

MethodPSNR↑SSIM↑T-ACC%↑T-NED↑P-ACC%↑P-NED↑
Bicubic19.70950.621142.900.697353.000.8093
NAFNet20.16850.639850.200.805361.500.9021
TSRN19.47080.615944.800.712951.100.7799
TBSRN19.77050.614746.200.731157.500.8344
TATT19.96470.618246.200.728554.800.8119
MARCONet18.88770.654848.700.754556.400.8260
DiffTSR18.19030.655748.500.748160.300.8466
TeReDiff18.44570.577542.400.685648.000.7566
TOLA teacher20.63140.671752.600.817467.100.9225
PixelSTR (ours)20.57860.669151.300.812566.400.9207

Photo variation. PixelSTR’s median photo PSNR is 20.27 dB (middle 50%, 19.11–21.92 dB). We average crops within each photo. The table gives PixelSTR minus bicubic, with equal weight for all 192 photos.

Per-photo differenceMean [95% CI]Median [Q1, Q3]
PSNR (dB)+0.8834 [+0.8373, +0.9336]+0.8891 [+0.6961, +1.1059]
SSIM+0.0443 [+0.0408, +0.0497]+0.0426 [+0.0338, +0.0577]
T-ACC (pp)+8.38 [+7.24, +9.31]-
P-ACC (pp)+13.18 [+12.21, +14.39]-

W3. Source quality. KD beats GT on PSNR in all nine training-defined strata. DiffTSR (§4.1) keeps 63,644/468,253 crops (≤24 characters, short side ≥64 px, aspect ratio >2), split 62,371/1,273 train/val. At P5/50/95, width is 180/423/1,324 px, height 66/102/319 px, length 2/4/12 characters, RMS contrast 0.0741/0.1838/0.3091, Laplacian variance 0.000727/0.009348/0.047351. Previews of 200 random source crops show blur, low contrast and occlusion. We degrade 512×128 HR with Real-ESRGAN or BSRGAN (50/50).

PropertyTertilenΔPSNR↑ΔLPIPS↓ΔT pp↑ΔP pp↑
HeightLow439+2.3777−0.1487+7.97+6.15
HeightMiddle398+2.1760−0.1732+6.53+3.77
HeightHigh436+2.2069−0.1930+10.09+4.59
Laplacian var.Low425+2.5882−0.1683+9.18+4.94
Laplacian var.Middle410+2.1667−0.1687+8.05+3.41
Laplacian var.High438+2.0177−0.1773+7.53+6.16
ContrastLow419+2.3347−0.1699+10.26+5.49
ContrastMiddle432+2.0767−0.1580+9.26+4.17
ContrastHigh422+2.3619−0.1871+5.21+4.98

Holdout gains (paired 95% CI). PSNR +2.2561 [2.1664,2.3488] dB, LPIPS −0.1715 [−0.1769,−0.1661], T +8.25 [6.60,9.90] pp, P +4.87 [3.38,6.36] pp.

W4. Severity. TOLA exceeds DiffTSR in Y-PSNR and both OCR accuracies in all six conditions. Under severe blur (σ=2.4), PixelSTR reaches 29.8633 dB and 94.50%/68.50% T/P. Under heavy JPEG compression (Q=10), it reaches 24.4244 dB and 64.00%/49.00%. Cells give Y-PSNR/T%/P% on 200 held-out images at ×4.

DegradationBicubicDiffTSRTOLAPixelSTR
Blur σ=0.625.7094/87.50/66.5028.9037/93.50/68.5031.0254/97.00/70.5029.6915/96.00/70.00
Blur σ=1.225.8437/88.50/67.0028.2114/92.00/67.5031.0829/96.00/70.0029.8982/96.00/70.00
Blur σ=2.425.5903/85.50/66.0026.9884/88.50/68.0030.9722/96.00/69.0029.8633/94.50/68.50
JPEG Q=9525.5042/88.00/66.5028.9365/94.50/67.0030.9313/95.00/69.0029.7454/96.00/68.00
JPEG Q=5024.1728/82.50/62.0027.0300/86.50/61.0029.1386/90.00/68.0027.8683/87.00/63.00
JPEG Q=1021.6309/59.50/29.5023.9745/63.50/40.5025.5378/67.50/50.0024.4244/64.00/49.00

Appendix I.1, Fig. 8, p. 21 compares DiffTSR, TOLA and PixelSTR on matched LR/HR crops under Gaussian blur (σ=2.4) and JPEG compression (Q=10). Fig. 9, p. 21 shows two severe-blur failures on dense text (a) and curved text (b). In both examples, both OCR recognizers misread the full string on all three models’ outputs.

References

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.