acceptodds
Under review as a conference paper at ICLR 2027

From Generation to Correction: Autoregressive Bitwise Correction for Image Super-Resolution

Abstract

Visual autoregressive (VAR) models have achieved remarkable progress in image generation, which has boosted their application to image super-resolution (ISR). However, existing VAR-based ISR methods tend to follow a generative paradigm conditioned on the low-quality (LQ) representation directly, without explicitly disentangling reliable content from degradation-induced discrepancies. This work focuses on bitwise VAR models that operate by quantizing latent features into binary codes. Within this binary latent space, we systematically analyze the bitwise behavior between LQ and high-quality (HQ) representations. We observe that LQ and HQ representations are highly similar at the bit level, motivating us to revisit ISR through an LQ-anchored bitwise correction formulation. This formulation aims to identify and correct LQ bits that disagree with their HQ counterparts while preserving consistent ones. This selective correction enables faithful restoration by leveraging reliable LQ information and concentrating the generative prior on degradation-induced discrepancies for high-frequency detail synthesis. Building on this insight, we propose Generative-Prior-enhanced Autoregressive Bitwise Correction (GPABC), a framework that autoregressively predicts LQ-relative flip masks by leveraging corrected representations from preceding scales as context alongside a pretrained VAR prior. To balance reconstruction fidelity and perceptual quality, we further introduce a scale-adaptive asymmetric focal flip loss. This loss penalizes false flips at coarse scales to preserve reliable LQ content, while penalizing missed flips at fine scales to recover high-frequency details with the generative prior. Extensive quantitative and qualitative experiments across four benchmarks show that GPABC outperforms state-of-the-art methods in perceptual quality while maintaining competitive reconstruction fidelity, validating the effectiveness of the proposed bitwise correction formulation for ISR.

Primary area:
generative models
Keywords:
Image Super-Resolution, Visual autoregressive models, Generative models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_K791OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper investigates the correspondence between low-quality (LQ) and high-quality (HQ) images in a binary latent space. Built upon the Infinity backbone, the proposed method predicts bit-flip probabilities relative to LQ bits and incorporates quantization-margin conditioning, shared-prefix modeling, curriculum learning, and a scale-adaptive asymmetric focal loss. The authors frame their approach as preserving reliable content while correcting degradation-induced discrepancies, and evaluate it on four image super-resolution benchmarks.

Critical strengths
  1. The analysis of LQ/HQ bit consistency across different scales, degradation levels, and quantization margins suggests that the quantization margin could serve as a useful conditioning signal. This is one of the most informative aspects of the paper.

  2. GPABC achieves the highest MANIQA and CLIPIQA scores across all four datasets and the best LPIPS score on DIV2K-Val. Therefore, it would be inaccurate to dismiss the method as entirely ineffective. However, the key question is whether these improvements genuinely substantiate the authors' claims regarding selective bit correction and fidelity preservation.

Critical weaknesses

1.The proposed bitwise correction framework appears to be largely a reformulation of conventional bit prediction rather than a fundamentally different learning paradigm. While the additional conditioning mechanisms and training strategies may provide practical benefits, their individual contributions are not sufficiently isolated. Consequently, it remains unclear whether the reported improvements genuinely arise from the proposed correction-based formulation or primarily from auxiliary design choices. This significantly weakens the paper's central novelty claim.

2.Although the method achieves strong results on several no-reference perceptual quality metrics, its performance on full-reference metrics, particularly on real-world benchmarks, is substantially worse than that of competing methods. This raises serious concerns about whether the proposed approach actually preserves reliable image content, as claimed, or simply favors perceptually appealing but potentially less faithful reconstructions. Moreover, the paper lacks direct evidence demonstrating that its correction mechanism effectively identifies and preserves reliable bits. As a result, the experimental findings do not convincingly support the claimed advantages in balancing perceptual quality and reconstruction fidelity.

Questions for authors

1.What is the fundamental methodological contribution of the proposed correction-based formulation? Given that the basic bit-flip prediction objective can be reformulated as conventional HQ-bit prediction under identical conditions, it is unclear whether the correction formulation itself introduces a meaningful algorithmic advantage. Could the authors clarify what fundamentally distinguishes GPABC from standard conditional bit prediction and provide controlled comparisons that isolate the contributions of the correction formulation, shared-prefix modeling, margin conditioning, and the proposed loss?

2.Does the proposed framework actually perform selective correction as claimed? The current analysis primarily characterizes statistical relationships between LQ and HQ representations, but does not directly establish that the trained model preserves reliable bits while selectively correcting erroneous ones. Could the authors provide quantitative evidence of actual correction behavior during autoregressive inference, including false-flip and missed-flip rates across scales, and demonstrate improvements over a matched direct bit-prediction baseline? Without such evidence, the central correction-based interpretation remains insufficiently validated.

3.How can the claimed fidelity-preservation and perceptual-quality advantages be reconciled with the substantially worse full-reference results on RealSR and DRealSR? GPABC underperforms competing methods across PSNR, SSIM, LPIPS, and DISTS on these benchmarks, despite achieving stronger no-reference quality scores. Could the authors demonstrate that GPABC achieves a genuinely better perception–distortion trade-off under matched evaluation conditions? Otherwise, the current results appear insufficient to support the claimed advantages in preserving reliable image content.

8 comments
AuthorsOpenReview ↗

Author Response – Part 1

We sincerely thank the reviewer for recognizing our informative bit-level analysis, and for providing insightful comments including the constructive feedback allowing us to improve our work.

W1 & Q1. Methodological Contributions and Component Ablations

We thank the reviewer for this insightful comment. Under identical conditioning and training configurations, the LQ-relative flip prediction objective and direct HQ-bit prediction are mathematically equivalent. Specifically, for a single bit, let bb be the binary LQ bit where sign(qLQ)=2b−1\text{sign}(q^{\text{LQ}}) = 2b - 1, yy the HQ target, and z=l1−l0z=l^1-l^0 the logit difference between the predicted positive (l1l^1) and negative (l0l^0) bit values. Following Equation (7) in the paper, the flip logit is formulated as sign(qLQ)(l0−l1)=(1−2b)z\text{sign}(q^{\text{LQ}})(l^0 - l^1) = (1-2b)z, and the target yy is mapped to the flip label b⊕yb\oplus y. This mapping preserves the BCE loss and its gradients.

Although mathematically equivalent, the correction perspective better characterizes the nature of image restoration, whose core goal, unlike image generation, is to preserve reliable content already present in the LQ input while correcting degradation-induced errors. This correction perspective directly inspired us to design LQ-anchored bit and margin conditioning. GPABC injects embeddings of the current-scale LQ bits and their quantization margins to preserve bits consistent with HQ while flipping inconsistent ones. In contrast, under the prediction perspective, existing autoregressive SR methods (e.g., VARSR, PURE, and DVAR) condition on the LQ image via LQ-derived [SOS] and context tokens, without taking the current-scale LQ information as explicit inputs. Table R1 compares this existing LQ conditioning with our LQ-anchored bit and margin conditioning on RealSR, with all other components and settings held fixed. As can be seen, our LQ-anchored conditioning improves both pixel-level fidelity and perceptual quality, e.g., increasing PSNR from 23.07 to 23.54 and MANIQA from 0.6389 to 0.6695. These results verify the advantage of explicit LQ-anchored bit and margin conditioning over existing LQ conditioning strategies.

To further demonstrate the effectiveness of each component, we conduct controlled ablation experiments isolating individual components. Table R1 shows that removing the quantization-margin embedding (w/o margin embedding) deprives the model of bit-reliability cues. The model may over-flip reliable bits under LSAF\mathcal{L}_{\mathrm{SAF}}, which slightly increases certain NR-IQA scores but leads to a degradation across all full-reference fidelity metrics. Removing LSAF\mathcal{L}_{\mathrm{SAF}} (w/o LSAF\mathcal{L}_{\mathrm{SAF}}) causes the model to tend toward conservative bit preservation, as unchanged bits dominate across scales and are easier to learn than bit flips. While this slightly improves reference-based metrics, it leads to a drop across all perceptual quality metrics. In addition, using independent encoding (w/o shared prefix) introduces accumulated context shifts across scales and leads to consistent degradation across all metrics, with PSNR dropping to 23.28 and CLIPIQA dropping to 0.7048.

Table R1. Ablations of GPABC’s conditioning and design components on RealSR. The variant without margin embedding retains the current-scale LQ-bit input. The best result in each column is bolded.

MethodPSNR ↑SSIM ↑LPIPS ↓DISTS ↓MANIQA ↑CLIPIQA ↑MUSIQ ↑TOPIQ ↑
w/o current-scale LQ input (Prediction Style)23.070.62400.40170.26280.63890.756170.560.7029
w/o margin embedding22.600.56290.45900.28820.70180.787370.290.7347
w/o LSAF\mathcal{L}_{\mathrm{SAF}}23.540.65270.37690.24890.63740.697269.980.6954
w/o shared prefix23.280.60720.42380.27820.56480.704868.560.6963
GPABC23.540.63960.40690.26880.66950.783170.660.7266

W2 & Q3. Content Preservation and Fidelity–Perception Balance

We appreciate the reviewer’s insightful comments. To address your concern that GPABC may favor perceptually appealing but less faithful reconstructions, we conducted a user study that evaluates fidelity to HQ as a separate criterion. We further examined the perception–distortion trade-off of GPABC under matched evaluation conditions, and we also provide direct evidence on the preservation of reliable bits.

AuthorsOpenReview ↗

Author Response – Part 2

User Study of Reconstruction Fidelity. We conducted a user study with 15 participants using 50 image groups randomly sampled from DIV2K-Val, RealSR, and DRealSR. Each group includes LQ, HQ, and the corresponding outputs from AR-based methods, DM-based methods, and our GPABC. All outputs were presented anonymously in randomized order. For each image group, all participants were asked to select the output with (i) the highest fidelity to HQ, (ii) the best visual quality, and (iii) the best overall balance between fidelity and perceptual quality, yielding 750 votes per criterion. As shown in Table R2, GPABC leads across all three evaluation dimensions with vote shares of 45.20%, 41.07%, and 47.60%, which demonstrates that the additional detail generated by GPABC is preferable and remains faithful to the input.

Table R2. Human comparison results.

CriterionStableSROSEDiffPiSA-SRVARSRPUREDVARNSARMGPABC
Fidelity3.60%6.00%11.07%5.33%6.93%6.13%15.73%45.20%
Quality3.20%3.07%6.67%4.93%12.80%12.27%16.00%41.07%
Overall3.20%4.53%9.33%4.80%7.33%6.00%17.20%47.60%

Analysis of FR-IQA Scores and Fidelity–Perception Balance. To understand the lower FR-IQA scores on RealSR and DRealSR, we evaluated the HQ reference images themselves using NR-IQA metrics. As shown in Table R3, the HQ reference images exhibit relatively low NR-IQA scores, particularly on RealSR and DRealSR, as they often retain mild optical blur or sensor noise. Consequently, when a generative restoration model removes these imperfections and synthesizes sharper details, its similarity to the HQ decreases. This partially explains the lower FR-IQA scores of GPABC on RealSR and DRealSR. We think it should be taken into consideration that HQ references are not necessarily ideal targets, as their imperfections may also contribute to the discrepancy between FR-IQA and NR-IQA metrics.

Table R3. NR-IQA scores of the HQ reference images.

DatasetMANIQA ↑CLIPIQA ↑MUSIQ ↑TOPIQ ↑
DIV2K-Val0.38840.620465.550.5776
RealSR0.32950.459057.460.4140
DRealSR0.31520.475150.560.3932

To further verify the perception–distortion trade-off under matched evaluation conditions, we trained a GPABC variant, GPABC-M, using milder synthetic degradation parameters so that the sharpness of its restored images aligns more closely with the HQ references. As shown in the table, GPABC-M reaches a PSNR of 25.45, which is comparable to the 25.50 of PiSA-SR, while achieving a higher MANIQA score of 0.5753 compared with 0.4810 of PiSA-SR. Moreover, compared with AR-based methods such as PURE and DVAR at comparable NR-IQA levels, GPABC-M achieves higher reconstruction fidelity across FR-IQA metrics, improving PSNR to 25.45 compared with 23.38 of PURE and 23.57 of DVAR. These comparisons support that our method achieves a favorable fidelity–perception balance.

Table R4. Comparison of all methods and GPABC-M on real-world RealSR benchmark.

MetricStableSRSeeSRSUPIROSEDiffPiSA-SRVARSRVARestorerPUREDVARNSARMGPABCGPABC-M
PSNR ↑24.8025.1522.9725.1525.5025.5924.9123.3823.5723.5423.5425.45
SSIM ↑0.71860.72100.55150.73410.74180.72660.69060.65020.67510.64120.63960.7357
LPIPS ↓0.29730.30080.48160.29200.26720.32250.27330.34960.32200.36940.40690.3056
DISTS ↓0.21590.22240.27170.21280.20440.23580.21880.23480.23150.25470.26880.2249
MANIQA ↑0.44220.54370.43350.47170.48100.54820.44370.44950.55560.60460.66950.5753
CLIPIQA ↑0.61990.67030.57420.66880.67040.69900.61480.66230.68230.73560.78310.7217
MUSIQ ↑66.0969.8261.0069.0970.1571.2368.5465.4270.3071.1370.6669.29
TOPIQ ↑0.58990.68880.56510.62480.63690.69900.61750.59260.69820.73150.72660.7001

Identification and Preservation of Reliable Bits. To directly verify that GPABC identifies and preserves reliable bits, Table R5 reports the preservation rate (the proportion of consistent LQ–HQ bits kept unchanged) across all 13 scales on RealSR. As can be seen, at coarse scales (s1–s4) encoding global structures and low-frequency content, GPABC preserves over 85% of reliable bits (up to 88.69% at s1 and 88.68% at s2), while moderately decreasing to 67.98%–79.58% at finer scales (s8–s13) as our scale-adaptive loss actively encourages bit flips to synthesize missing high-frequency details.

Table R5. Per-scale preservation rates on RealSR.

Metrics1s2s3s4s5s6s7s8s9s10s11s12s13
Preservation rate ↑0.88690.88680.87200.85090.84160.83370.82820.79580.77100.75570.71510.67980.6975
AuthorsOpenReview ↗

Author Response – Part 3

Q2. Selective Correction Behavior During Autoregressive Inference

We thank the reviewer for this insightful comment. Following your suggestion, we evaluated the actual bitwise correction behavior of GPABC during autoregressive inference on the real-world RealSR benchmark and compared it against the matched prediction-style baseline. At each scale kk, we measure three metrics. Bit agreement rate measures the proportion of predicted bits that match the target HQ bits. False-flip rate measures the proportion of bits that the model erroneously flips among reliable LQ bits that agree with HQ. Missed-flip rate measures the proportion of bits that the model fails to flip among erroneous LQ bits that disagree with HQ. Table R6 reports these metrics across all scales, revealing three findings that directly validate our selective correction claim:

  • GPABC achieves higher bit agreement rates across all scales: By explicitly conditioning on current-scale LQ bits and quantization margins as correction anchors, GPABC accurately aligns degraded representations with their HQ targets and achieves higher bit agreement rates across all 13 scales from s1 to s13, for instance improving s1 from 0.8387 to 0.8784, s2 from 0.7891 to 0.8647, s8 from 0.6617 to 0.7342, and s13 from 0.5978 to 0.6315.

  • GPABC preserves reliable bits with lower false-flip rates across all scales: Across all scales, especially coarse scales that encode global structures and low-frequency content, GPABC achieves lower false-flip rates than the prediction-style baseline. For example, GPABC reduces the false-flip rate from 0.1623 to 0.1131 at s1, from 0.2055 to 0.1132 at s2, and from 0.3741 to 0.3025 at s13. This demonstrates that GPABC effectively identifies and preserves reliable LQ bits, retaining over 88.6% of reliable bits at s1 and s2 and over 82.8% across scales s1 to s7.

  • GPABC exhibits a relatively higher missed-flip rate across all scales: Without current-scale LQ anchors, the prediction-style baseline flips bits more aggressively, resulting in a lower missed-flip rate, such as 0.3743 vs. 0.4150 at s7, but at the cost of a substantially higher false-flip rate, such as 0.2607 vs. 0.1718, consequently leading to a lower bit agreement rate of 0.7103 vs. 0.7691. In contrast, GPABC performs selective correction, which ultimately yields superior overall bit agreement rates across every scale.

Table R6. Per-scale correction behavior during autoregressive inference on RealSR.

MetricMethods1s2s3s4s5s6s7s8s9s10s11s12s13
Bit agreement rate ↑Baseline0.83870.78910.77510.73880.72340.71160.71030.66170.64570.64230.60190.59270.5978
Bit agreement rate ↑GPABC0.87840.86470.84240.81040.79700.78240.76910.73420.70950.68680.65360.62590.6315
False-flip rate ↓Baseline0.16230.20550.21290.24770.26050.26920.26070.31810.33270.32600.37710.38650.3741
False-flip rate ↓GPABC0.11310.11320.12800.14910.15840.16630.17180.20420.22900.24430.28490.32020.3025
Missed-flip rate ↓Baseline0.14430.23230.26820.30420.32550.34490.37430.39070.40870.43470.44720.45510.4635
Missed-flip rate ↓GPABC0.15930.23480.27930.33960.35940.38610.41500.43630.44090.47150.47310.47600.4892

We will add this per-scale correction behavior analysis during autoregressive inference on RealSR to the appendix of the revised manuscript.

Reviewer_deM5OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes an image SR method named GPABC by reformulating ISR as LQ-anchored bitwise correction in the discrete latent space of a bitwise VAR model. Experiments show strong NR and competitive FR scores, with small trainable parameters and fast inference time.

Critical strengths
  1. Reformulating ISR as LQ-anchored bitwise correction is novel. Freezing Infinity-2B and training only LoRA adapters and small projections keeps trainable parameters low while leveraging a strong pretrained prior.
  2. Well-thought method design. Shared-prefix bit/margin computation, LQ-relative flip prediction, scale-adaptive asymmetric focal loss, etc. are logically connected.
  3. Ablation studies on scale weights, loss, curriculum, and encoding strategy. Inference time and parameter comparisons are also provided. The implementation details, hyperparameters, and promises of code/checkpoint release facilitate reproducibility.
Critical weaknesses
  1. The method builds on Infinity and existing VAR-based ISR. Broader generalizability is untested. It is not clear whether they transfer to other VAR models, VQ tokenizers, or broader degradation distributions.
  2. The proposed method achieves top NR IQA scores but not FR metrics. Some diffusion-based methods still have higher PSNR/SSIM. There are more dedicated approaches for quality assessment of image SR, that are highly relevant but not mentioned or used in evaluation.
  3. Ablation scope is still limited. Main ablations are on DIV2K-Val only. No real-world benchmarks. Sensitivity to key hyperparameters and the curriculum schedule is not deeply investigated.
Questions for authors
  1. During inference how do you handle error accumulation? Can you provide per-scale flip accuracy or error propagation analysis?
  2. How does it generalize to arbitrary scale factors and output resolutions, and other VAR/tokenizer backbones?
  3. Include more dedicated IQA methods for ISR in the evaluation, for example,
  • W. Zhou et al, "Quality assessment of image super-resolution: balancing deterministic and statistical fidelity," ACM Multimedia 2022.
  • W. Zhou et al, "image super-resolution quality assessment: structural fidelity versus statistical naturalness," QoMEX 2021.
  1. How sensitive is performance to gamma, alpha, lambda, and the curriculum rollout probability? Are these robust across benchmarks?
3 comments
AuthorsOpenReview ↗

Author Response – Part 1

We sincerely thank the reviewer for recognizing the novelty of reformulating ISR as LQ-anchored bitwise correction, and for providing insightful comments including the constructive feedback allowing us to improve our work.

W1 & Q2. Scope and Generalization of Our Method

We appreciate the reviewer’s insightful comments. Below, we clarify the scope and generalization capability of GPABC across different VAR backbones/tokenizers, degradation distributions, and scale factors/output resolutions.

  • VAR models and tokenizers. Implemented on Infinity and its BSQ tokenizer, GPABC is specifically designed for binary quantized VAR models, where deciding whether to flip a bit uniquely determines its corrected value. Consequently, it does not directly transfer to non-binary tokenizers such as categorical VQ codebooks, which require predicting both whether to change a codeword and which replacement codeword to select. We will include this limitation in the revised manuscript and will explore more general correction mechanisms across diverse VAR tokenizers in future work.

  • Broader degradation distributions. During training, our LQ images are synthesized using the Real-ESRGAN degradation pipeline, whereas our test benchmarks include real-world RealSR, DRealSR, and RP60 datasets. Their LQ degradations arise from real image acquisition and processing, including scene- and device-dependent blur, noise, and artifacts. The results reported in Table 1 of the main paper demonstrate that our method generalizes beyond synthetic training degradations to diverse real-world degradation distributions.

  • Scale factors & Output resolutions. Like the DM- and AR-based baselines in Table 1, GPABC resizes the input LQ image to the target resolution before encoding, making the model agnostic to specific upscaling factors and naturally capable of arbitrary-scale SR. However, GPABC does not support arbitrary output resolutions: its output size is fixed by Infinity’s tokenizer grid and scale schedule, and altering it requires reconfiguring the schedule and retraining.

We will add a discussion of the limitations regarding transferring to other VAR representation types and generalizing to arbitrary output resolutions in Section 6 (Limitation and Future work) of the revised manuscript. Future work will explore more general models across diverse token representations as well as image restoration at arbitrary output resolutions.

W2 & Q3. Fidelity–Perception Balance and Additional Quality Assessment Metrics

We thank the reviewer for this valuable suggestion. Following your suggestion, we evaluated the models using SFSN [QoMEX'21] and tested the applicability of SRIF [ACMMM'22].

Results on SFSN. As shown in Table R1, our proposed GPABC achieves the best SFSN scores across all three test sets. Since SFSN jointly evaluates Structural Fidelity (SF) and Statistical Naturalness (SN), this superior performance strongly validates that GPABC achieves a favorable balance between fidelity and perceptual quality.

Table R1. SFSN results on DIV2K, DRealSR, and RealSR.

DatasetGPABCPiSA-SRDVARNSARMOSEDiffPUREStableSRVARSR
DIV2K1.05431.00250.96210.99340.98771.01361.00350.9919
DRealSR1.03271.00920.98621.01550.99710.99091.00511.0248
RealSR1.05171.03250.99441.01431.02091.03341.02531.0350

Applicability Analysis of SRIF. We also attempted to evaluate our method using the official SRIF code, but it cannot be faithfully reproduced on our benchmarks due to two limitations:

  • Out-of-Range Mapping: SRIF maps features to the MOS scale via a logistic function fitted on the Waterloo Database. On our test sets, 17.9% of the images exceed this fitting range. Unclipped extrapolation causes severe outliers, while clipping collapses distinct samples to the same boundary, erasing quality differences. Both introduce uncontrollable biases.
  • Lack of MOS for Re-fitting: Re-fitting this function to our data requires ground-truth MOS, which all three benchmarks lack.
AuthorsOpenReview ↗

Author Response – Part 2

Human Comparisons. We also conducted a user study with 15 participants using 50 image groups randomly sampled from DIV2K-Val, RealSR, and DRealSR. Each group includes LQ, HQ, and the corresponding outputs from AR-based methods (VARSR, PURE, DVAR, and NSARM), DM-based methods (StableSR, OSEDiff, and PiSA-SR), and our GPABC. All outputs were presented anonymously in randomized order. For each image group, all participants were asked to select the output with (i) the highest fidelity to HQ, (ii) the best visual quality, and (iii) the best overall balance between fidelity and perceptual quality, yielding 750 votes per criterion. As shown in Table R2, GPABC leads across all three evaluation dimensions with vote shares of 45.20%, 41.07%, and 47.60%, which further confirms its fidelity–perception balance.

Table R2. Human comparison results.

CriterionStableSROSEDiffPiSA-SRVARSRPUREDVARNSARMGPABC
Fidelity3.60%6.00%11.07%5.33%6.93%6.13%15.73%45.20%
Quality3.20%3.07%6.67%4.93%12.80%12.27%16.00%41.07%
Overall3.20%4.53%9.33%4.80%7.33%6.00%17.20%47.60%

In summary, both the objective dedicated metric (SFSN) and human comparison results consistently favor GPABC, confirming that it achieves a favorable balance between reconstruction fidelity and perceptual quality. We will cite both SFSN and SRIF methods in Sec. 5.1 and include the SFSN and user study results in Sec. 5.2 of the revised manuscript.

W3 & Q4. Real-World Ablations and Hyperparameter Robustness

We thank the reviewer for this constructive suggestion. Following your suggestion, we conducted component ablation studies and hyperparameter sensitivity analyses on the real-world RealSR benchmark.

Ablations on the real-world RealSR benchmark. We conduct ablation experiments on the real-world RealSR benchmark. The results shown in Table R3 validate the effectiveness of each component in GPABC.

  • Fixed ρ=1\rho=1 keeps the LQ conditioning weight unattenuated at fine scales, improving FR metrics with LPIPS decreasing from 0.4069 to 0.3671 at the expense of detail synthesis and NR-IQA scores with CLIPIQA dropping from 0.7831 to 0.7775.
  • w/o LSAF\mathcal{L}_{\mathrm{SAF}} removes the scale-adaptive asymmetric focal loss. Since unchanged bits dominate across scales and are easier to learn than bit flips, the model tends toward conservative bit preservation, improving FR metrics with LPIPS decreasing from 0.4069 to 0.3769 at the cost of a severe drop in perceptual quality with CLIPIQA dropping from 0.7831 to 0.6972.
  • w/o curriculum, which corresponds to full teacher-forcing, suffers from exposure bias during autoregressive inference and worsens fidelity and perceptual metrics, for instance LPIPS worsens to 0.4157 and CLIPIQA drops to 0.7372.
  • w/o shared prefix, which uses independent encoding, introduces accumulated context shifts across scales and leads to consistent degradation across all metrics, with LPIPS worsening from 0.4069 to 0.4238 and CLIPIQA dropping from 0.7831 to 0.7048.

Table R3. Ablation studies on different components of GPABC on real-world RealSR benchmark.

MethodLPIPS ↓CLIPIQA ↑MUSIQ ↑TOPIQ ↑
Fixed ρ=1\rho=10.36710.777569.120.7103
w/o LSAF\mathcal{L}_{\mathrm{SAF}}0.37690.697269.980.6954
w/o curriculum0.41570.737269.120.7077
w/o shared prefix0.42380.704868.560.6963
GPABC0.40690.783170.660.7266

Hyperparameter sensitivity. We further evaluate the sensitivity of hyperparameters α\alpha, γ\gamma, and λ\lambda on RealSR by varying one parameter at a time with all other settings fixed. As shown in Table R4, GPABC remains stable across a reasonable range of hyperparameter variations. Moreover, we use the exact same hyperparameter configuration across all evaluated datasets without dataset-specific tuning. Under this single shared configuration, GPABC achieves SoTA performance, demonstrating strong cross-dataset robustness.

  • Varying α\alpha: A smaller α=1\alpha=1 weakens the scale-wise asymmetry, lowering both fidelity and perceptual metrics compared to α=2\alpha=2, which is the default setting used in GPABC. Increasing α\alpha to 44 degrades both fidelity and perceptual quality metrics, whereas the default setting achieves the best overall balance.
AuthorsOpenReview ↗

Author Response – Part 3

  • Varying γ\gamma: A large γ\gamma weakens the fine-scale flip encouragement of LSAF\mathcal{L}_{\mathrm{SAF}}, preserving more LQ bits and shifting toward fidelity. Thus, γ=4\gamma=4 improves LPIPS over γ=2\gamma=2 (0.3726 vs. 0.4069) but lowers NR-IQA (CLIPIQA drops from 0.7831 to 0.7663). We use γ=2\gamma=2 by default for a balanced trade-off.

  • Varying λ\lambda: Decreasing the loss weight to λ=0.5\lambda=0.5 favors fidelity, yielding LPIPS of 0.3784 with reduced perceptual quality, while increasing it to λ=2\lambda=2 pushes the model toward higher CLIPIQA of 0.7889 at the cost of lower fidelity with LPIPS of 0.4448. We set λ=1\lambda=1 by default to better balance fidelity and perceptual quality.

Table R4. Sensitivity to hyperparameters on real-world RealSR benchmark.

ConfigurationLPIPS ↓CLIPIQA ↑MUSIQ ↑TOPIQ ↑
α=1\alpha=10.41080.774469.730.7173
α=2\alpha=20.40690.783170.660.7266
α=4\alpha=40.42090.769870.040.7161
γ=1\gamma=10.42370.782169.890.7130
γ=2\gamma=20.40690.783170.660.7266
γ=4\gamma=40.37260.766368.770.7059
λ=0.5\lambda=0.50.37840.767068.890.7105
λ=1\lambda=10.40690.783170.660.7266
λ=2\lambda=20.44480.788969.980.7230

Curriculum rollout probability. This probability controls the replacement of HQ bits with predicted bits at disagreement positions. By default, we apply teacher forcing for the first 2K steps, then linearly increase the replacement probability to 11 over the remaining steps. We also tested linearly increasing the replacement probability to 0.250.25, 0.50.5, and 0.750.75 during training under identical settings. As shown in Table R5, among the tested probabilities, GPABC achieves the best performance across all metrics. This supports increasing the probability to full predicted-bit replacement at disagreement positions by the end of training.

Table R5. Sensitivity to the curriculum rollout probability on real-world RealSR benchmark.

ProbabilityLPIPS ↓CLIPIQA ↑MUSIQ ↑TOPIQ ↑
0.250.41500.769570.360.7166
0.500.43500.771269.310.7065
0.750.42920.776369.090.7040
1.00 (GPABC)0.40690.783170.660.7266

We will include these RealSR ablations and sensitivity analyses in the appendix of the revised manuscript.

Q1. Error Accumulation and Per-Scale Correction Analysis during Inference

We thank the reviewer for this insightful question. We address inference-time error accumulation through progressive prefix curriculum learning. The first 2K training steps use HQ prefixes to allow the model to learn bit flip prediction under a clean multi-scale context. During the remaining 20K steps, detached predicted bits replace HQ bits at disagreement positions with a probability that increases linearly from 0 to 1. LQ bits, margins, and HQ targets are recomputed under the resulting prefix. This trains the model to make corrections under imperfect histories and mitigates exposure bias during autoregressive inference. Table R3 shows that curriculum learning consistently improves both reconstruction fidelity and perceptual quality over full teacher forcing, denoted as w/o curriculum, improving LPIPS from 0.4157 to 0.4069, CLIPIQA from 0.7372 to 0.7831, MUSIQ from 69.12 to 70.66, and TOPIQ from 0.7077 to 0.7266.

Furthermore, we compare the per-scale flip accuracy between full teacher forcing and our method in Table R6 to analyze error propagation. During inference, each model constructs its own prefix from its predicted corrected bits at preceding scales. At each scale, LQ and HQ bits are computed under that shared prefix, and the predicted flip mask is compared with mask label. Table R6 shows that as the scale index increases, the bit accuracy of the full teacher-forcing model drops from 0.8509 at s1 to 0.5526 at s13, demonstrating severe cross-scale error accumulation as early prediction errors propagate into finer-scale prefixes. In contrast, GPABC maintains consistently higher flip accuracy across all 13 scales, and its advantage over full teacher forcing is generally larger at finer scales, improving accuracy from 0.7036 to 0.7691 at s7 and from 0.5526 to 0.6315 at s13. In summary, these results demonstrate that our curriculum learning strategy effectively mitigates error accumulation.

Table R6. Per-scale flip accuracy during autoregressive inference on RealSR.

Methods1s2s3s4s5s6s7s8s9s10s11s12s13
w/o curriculum0.85090.81740.79370.76090.73850.72350.70360.66350.63820.61120.58700.57680.5526
GPABC0.87840.86470.84240.81040.79700.78240.76910.73420.70950.68680.65360.62590.6315

We will add the per-scale flip accuracy curves during autoregressive inference on RealSR to the appendix of the revised manuscript.

Reviewer_gKMDOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes GPABC (Generative-Prior-enhanced Autoregressive Bitwise Correction) which reframes ISR as an LQ-anchored bitwise correction task rather than conditional generation. GPABC builds on a frozen Infinity-2B backbone with LoRA adapters which corrects LQ bits toward their HQ counterparts through LQ-relative flip-mask prediction. A scale-adaptive asymmetric focal flip loss is also proposed to preserve reliable LQ content and synthesize high-frequency details. Extensive expreiments show GPABC achieves SOTA on multiple metrics and datasets.

Critical strengths
  1. The shift from "conditional generation" to "bitwise correction" is conceptually original, and the bit-level analysis provides concrete, measurable support for the motivation.
  2. The idea of shared-prefix encoding is reasonable and effective. Comparing residual bits generated under different histories can conflate degradation effects with differences accumulated at preceding scales. Using a common prefix is a technically well-motivated way to reduce this ambiguity. The benefits of this component are supported experimentally.
  3. The training design is supported by targeted ablations and requires relatively few trainable parameters. The ablations support the usefulness of both the curriculum and the additional loss within the proposed system. Adapting a frozen Infinity backbone through LoRA also provides a practical parameter-efficient training strategy.
  4. The no-reference quality improvements are consistent across multiple benchmarks. Consistency across a synthetic benchmark and three real-world benchmarks makes the no-reference improvements more convincing than a gain confined to one dataset. The comparison includes both diffusion-based and autoregressive alternatives, providing useful context.
Critical weaknesses

Please identify one or two critical weaknesses of the paper that support rejection. For each, explain briefly why it is consequential to your overall assessment. The goal is not to enumerate every negative point, but to identify and substantiate the considerations that most strongly determine your recommendation. Enter “None” if no material weakness can be identified.

  1. The claimed fidelity-perception balance is overstated on the paired real-world benchmarks. Section 5.2 describes GPABC as "maintaining comparable full-reference metric scores." However, Table 1 shows that GPABC lags behind on FR-IQA. Human comparisons and fidelity-matched operating points are needed to assess whether the additional detail is preferable and remains faithful to the input.
  2. The motivating statistics do not yet establish the claimed mechanism under the actual correction protocol. Section 3 does not clearly specify whether its statistics use independent prefixes or the shared-prefix construction adopted by GPABC. Moreover, cross-degradation overlap demonstrates dependence, not necessarily predictability from LQ observations. Real-pair and rollout-prefix analyses would strengthen the connection to the deployed method.
  3. Important data and inference details remain insufficiently specified for reproducibility and comparison. The paper does not specify the training-image sources, filtering criteria, or benchmark-overlap checks. It is also unclear how test-time text is obtained.
Questions for authors

Q1. What improvement is attributable specifically to correction rather than to conditioning and loss design? Could you compare direct HQ-bit prediction and LQ-relative flip prediction with identical shared prefixes, LQ/margin embeddings, LoRA capacity, curriculum, and scale weights, first without L_SAF and then with its equivalent bit-label formulation? Please clarify which operation is expected to make their behavior differ. This comparison would separate a useful interpretation of existing logits from changes in the learning problem, and would help locate the paper's main technical contribution precisely. Q2. How should the real-world fidelity losses be reconciled with the claim of faithful restoration? Human comparisons and fidelity-matched operating points are needed to assess whether the additional detail is preferable and remains faithful to the input. Q5. Can the data, text-conditioning, and postprocessing protocol be made fully reproducible?

4 comments
AuthorsOpenReview ↗

Author Response – Part 1

We appreciate the reviewer for recognizing the originality of our bitwise correction formulation and bit-level analysis, as well as the constructive comments.

W1 & Q2. Fidelity–Perception Balance and Faithful Restoration on Paired Real-World Benchmarks

We appreciate the reviewer’s insightful comments. Following your suggestion, we conducted human comparisons and evaluated GPABC at a fidelity-matched operating point. We will also revise the wording to avoid overstating the full-reference results on paired real-world benchmarks.

Human Comparisons. We conducted a user study with 15 participants using 50 image groups randomly sampled from DIV2K-Val, RealSR, and DRealSR. Each group includes LQ, HQ, and the corresponding outputs from AR-based methods (VARSR, PURE, DVAR, and NSARM), DM-based methods (StableSR, OSEDiff, and PiSA-SR), and our GPABC. All outputs were presented anonymously in randomized order. For each image group, all participants were asked to select the output with (i) the highest fidelity to HQ, (ii) the best visual quality, and (iii) the best overall balance between fidelity and perceptual quality, yielding 750 votes per criterion. As shown in Table R1, GPABC leads across all three evaluation dimensions with vote shares of 45.20%, 41.07%, and 47.60%, which demonstrates that the additional detail generated by GPABC is preferable and remains faithful to the input.

Table R1. Human comparison results.

CriterionStableSROSEDiffPiSA-SRVARSRPUREDVARNSARMGPABC
Fidelity3.60%6.00%11.07%5.33%6.93%6.13%15.73%45.20%
Quality3.20%3.07%6.67%4.93%12.80%12.27%16.00%41.07%
Overall3.20%4.53%9.33%4.80%7.33%6.00%17.20%47.60%

Analysis of FR-IQA Scores and Fidelity–Perception Balance. To understand the lower FR-IQA scores on RealSR and DRealSR, we evaluated the HQ reference images themselves using NR-IQA metrics. As shown in Table R2, the HQ reference images exhibit relatively low NR-IQA scores, particularly on the RealSR and DRealSR benchmarks, as they often retain mild optical blur or sensor noise. Consequently, when a generative restoration model removes these imperfections and synthesizes sharper details, its similarity to the HQ decreases. This partially explains the lower FR-IQA scores of GPABC on RealSR and DRealSR. We think it should be taken into consideration that HQ references are not necessarily ideal targets, as their imperfections may also contribute to the discrepancy between FR-IQA and NR-IQA metrics.

Table R2. NR-IQA scores of the HQ reference images.

DatasetMANIQA ↑CLIPIQA ↑MUSIQ ↑TOPIQ ↑
DIV2K-Val0.38840.620465.550.5776
RealSR0.32950.459057.460.4140
DRealSR0.31520.475150.560.3932

Furthermore, to evaluate GPABC at a fidelity-matched operating point, we trained a GPABC variant, GPABC-M, using milder synthetic degradation parameters so that the sharpness of its restored images aligns more closely with the HQ references. As shown in the table, GPABC-M reaches a PSNR of 25.45, which is comparable to the 25.50 of PiSA-SR, while achieving a higher MANIQA score of 0.5753 compared with 0.4810 of PiSA-SR. Moreover, compared with AR-based methods such as PURE and DVAR at comparable NR-IQA levels, GPABC-M achieves higher reconstruction fidelity across FR-IQA metrics, improving PSNR to 25.45 compared with 23.38 of PURE and 23.57 of DVAR. These comparison results support that our method achieves a favorable fidelity–perception balance.

Table R3. Comparison of all methods and GPABC-M on real-world RealSR benchmark.

MetricStableSRSeeSRSUPIROSEDiffPiSA-SRVARSRVARestorerPUREDVARNSARMGPABCGPABC-M
PSNR ↑24.8025.1522.9725.1525.5025.5924.9123.3823.5723.5423.5425.45
SSIM ↑0.71860.72100.55150.73410.74180.72660.69060.65020.67510.64120.63960.7357
LPIPS ↓0.29730.30080.48160.29200.26720.32250.27330.34960.32200.36940.40690.3056
DISTS ↓0.21590.22240.27170.21280.20440.23580.21880.23480.23150.25470.26880.2249
MANIQA ↑0.44220.54370.43350.47170.48100.54820.44370.44950.55560.60460.66950.5753
CLIPIQA ↑0.61990.67030.57420.66880.67040.69900.61480.66230.68230.73560.78310.7217
MUSIQ ↑66.0969.8261.0069.0970.1571.2368.5465.4270.3071.1370.6669.29
TOPIQ ↑0.58990.68880.56510.62480.63690.69900.61750.59260.69820.73150.72660.7001

In our revised manuscript, we will refine the wording, add the user study results in Section 5.2, and include the comparison results at the fidelity-matched operating point in the appendix to better demonstrate our fidelity–perception balance.

AuthorsOpenReview ↗

Author Response – Part 2

W2. Motivating Statistics under Real-Pair and Shared Rollout-Prefix Protocols

We thank the reviewer for this insightful observation. We clarify that the statistics in Section 3 of the submitted manuscript used independent prefixes. To verify the motivating statistics under the actual correction protocol deployed in GPABC, we performed the rollout-prefix analyses on 100 real-world LQ–HQ pairs from RealSR, alongside independent prefixes for comparison.

Scale-wise bit agreement under different degradation levels. Table R4 shows that LQ–HQ agreement remains higher at coarse scales and decreases with stronger degradation under shared rollout prefixes. For example, at scale s1, the agreement decreases from 0.9034 (2x) to 0.8572 (3x) and 0.8309 (4x), and for 4x degradation it gradually decreases from 0.8309 at s1 to 0.6409 at s13. Moreover, shared prefixes yield higher mean agreement than independent prefixes at every degradation level (e.g., increasing the 4x mean from 0.7102 to 0.7436). We believe this is because shared rollout prefixes compute LQ and HQ residuals under the same cumulative context, which avoids accumulating quantization differences from earlier scales. Overall, these results confirm that the scale-wise bit agreement pattern holds under shared rollout prefixes.

Table R4. Scale-wise LQ–HQ bit agreement ratio under different degradation levels on RealSR.

LQ degradationPrefixs1s2s3s4s5s6s7s8s9s10s11s12s13Mean
2xIndependent0.90340.81800.81020.79070.79360.78830.80580.75040.75090.75270.69240.67420.69970.7716
2xShared rollout0.90340.90020.88880.87480.86540.85750.84860.83260.81220.79590.77920.76000.74540.8357
3xIndependent0.85720.75630.76440.75390.75880.75460.76870.71530.71070.70320.64960.63000.63920.7278
3xShared rollout0.85720.84840.83580.81980.80680.79800.78690.76490.74090.72330.70150.68130.66830.7718
4xIndependent0.83090.73290.74580.73990.74530.74120.75460.70210.69460.68320.63340.61280.61570.7102
4xShared rollout0.83090.81840.80940.79150.78060.76990.75930.73640.71110.69450.67180.65250.64090.7436

Directed inclusion of flip sets across degradation levels. Table R5 shows directed inclusion ratios between mismatch sets across degradation levels under shared rollout prefixes. Compared with independent prefixes, shared prefixes increase the mean directed inclusion ratio from 0.6353 to 0.6677 for 2x$\to$3x and from 0.6681 to 0.7086 for 3x$\to$4x. These results preserve the cross-degradation dependence observed with independent prefixes. Furthermore, we acknowledge that cross-degradation overlap demonstrates statistical dependence across degradation levels rather than direct learnability or predictability from LQ observations, and we will revise the corresponding wording in Section 3 accordingly.

Table R5. Directed inclusion ratios across degradation levels on RealSR.

Directed inclusionPrefixs1s2s3s4s5s6s7s8s9s10s11s12s13Mean
∣S2∩S3∣/∣S2∣\lvert S_2\cap S_3\rvert/\lvert S_2\rvertIndependent0.94750.80560.70930.65210.61420.59270.59130.56430.56590.57190.54590.54150.55710.6353
∣S2∩S3∣/∣S2∣\lvert S_2\cap S_3\rvert/\lvert S_2\rvertShared rollout0.94750.87770.80920.73590.66700.61660.59920.58750.57840.56840.58000.56540.54790.6677
∣S3∩S4∣/∣S3∣\lvert S_3\cap S_4\rvert/\lvert S_3\rvertIndependent0.98240.86190.77930.71150.65840.62790.61950.58860.58740.59040.55930.55280.56630.6681
∣S3∩S4∣/∣S3∣\lvert S_3\cap S_4\rvert/\lvert S_3\rvertShared rollout0.98240.91450.83990.77580.70730.66690.64540.63360.62520.60930.62210.60640.58280.7086
AuthorsOpenReview ↗

Author Response – Part 3

Relationship between bit margin and bit agreement. Table R6 shows that larger LQ margins consistently correspond to higher agreement under both protocols. For 4x with shared prefixes, agreement increases from 0.5412 in the lowest-margin bin to 0.9400 in the highest. Shared prefixes yield higher agreement in every reported bin, supporting LQ margin as a reliability cue under the actual shared-prefix construction.

Table R6. Relationship between LQ bit margin and LQ–HQ agreement ratio on RealSR.

LQ degradationPrefix[0, 0.025)[0.025, 0.050)[0.050, 0.100)[0.100, 0.200)[0.200, 0.400)≥0.400
2xIndependent0.54990.61560.71040.82230.91690.9706
2xShared rollout0.57130.67640.79190.90160.96380.9860
3xIndependent0.52840.58350.65950.76710.87480.9461
3xShared rollout0.54210.62230.71320.82070.91150.9615
4xIndependent0.52890.57500.64260.74410.85270.9292
4xShared rollout0.54120.60640.68470.78240.87840.9400

Channel-wise degradation sensitivity. We will add an analysis of channel-wise degradation sensitivity in the form of heatmaps to the appendix of the revised manuscript, showing that reliability remains heterogeneous across bit channels under shared rollout prefixes. Compared with independent prefixes, agreement is generally higher, while channel-dependent differences remain visible.

In summary, these results show that the motivating analytical patterns persist on real images under shared rollout prefixes. We will clarify the original prefix protocol in Section 3 and add these results in the form of more intuitive figures to the revised manuscript.

W3 & Q5. Reproducibility of Data, Text-Conditioning, and Postprocessing

We appreciate the reviewer’s insightful comments. To make our experimental pipeline fully reproducible, we clarify the training data sources, filtering criteria, benchmark-overlap checks, test-time text conditioning, and postprocessing protocol.

  • Training Data & Overlap Checks: Our training images are entirely sourced from LSDIR, the first 10k images of FFHQ, and 1M images collected from Visual China Group (VCG), a public stock photography platform. Specifically, images with a shorter side below 1024 are resized along the shorter side to 1024 before being center-cropped to 1024×10241024 \times 1024, while all others are directly center-cropped. Then, we filter the dataset to retain only images with scores above 80 as assessed by our in-house no-reference image quality assessment method. Strict filename and perceptual-hash deduplication against all evaluation benchmarks are performed to avoid data leakage. We will release the IDs of all training images upon acceptance to ensure full reproducibility.
  • Test-Time Text Conditioning: At test time, text conditioning is generated from the input LQ image via LLaVA-v1.5-13B with the fixed prompt: "Describe this image and its style in a very detailed manner."
  • Postprocessing Protocol: As noted in the submitted manuscript, following NSARM, we apply standard wavelet-based and AdaIN color alignment using the bicubic-upsampled LQ image as the color reference to prevent global color shifts.

We will explicitly describe these data preparation steps and the test-time captioning procedure in the appendix of the revised manuscript.

AuthorsOpenReview ↗

Author Response – Part 4

Q1. Correction formulation and technical contributions

We thank the reviewer for this insightful comment. Under the matched conditions you describe, direct HQ-bit prediction and LQ-relative flip prediction are mathematically equivalent. Specifically, for a single bit, let bb be the binary LQ bit where sign(qLQ)=2b−1\text{sign}(q^{\text{LQ}}) = 2b - 1, yy the HQ target, and z=l1−l0z=l^1-l^0 the logit difference between the predicted positive (l1l^1) and negative (l0l^0) bit values. Following Equation (7) in the paper, the flip logit is formulated as sign(qLQ)(l0−l1)=(1−2b)z\text{sign}(q^{\text{LQ}})(l^0 - l^1) = (1-2b)z, and the target yy is mapped to the flip label b⊕yb\oplus y. This mapping preserves the BCE loss and its gradients, and the equivalence also holds with LSAF\mathcal{L}_{\mathrm{SAF}}.

Although mathematically equivalent, the correction perspective better characterizes the nature of image restoration, whose core goal, unlike image generation, is to preserve reliable content already present in the LQ input while correcting degradation-induced errors. This correction perspective directly inspired us to design LQ-anchored bit and margin conditioning. GPABC injects embeddings of the current-scale LQ bits and their quantization margins to preserve bits consistent with HQ while flipping inconsistent ones. In contrast, under the prediction perspective, existing autoregressive SR methods (e.g., VARSR, PURE, and DVAR) condition on the LQ image via LQ-derived [SOS] and context tokens, without taking the current-scale LQ information as explicit inputs. Table R7 compares this existing LQ conditioning with our LQ-anchored bit and margin conditioning on RealSR, with all other components and settings held fixed. As shown in Table R7, our LQ-anchored conditioning improves both pixel-level fidelity and perceptual quality, e.g., increasing PSNR from 23.07 to 23.54 and MANIQA from 0.6389 to 0.6695. These results verify the advantage of explicit LQ-anchored bit and margin conditioning over existing LQ conditioning strategies. Beyond LQ-anchored conditioning, our scale-adaptive asymmetric focal loss (LSAF\mathcal{L}_{\mathrm{SAF}}) is also inspired by this correction framework. Moreover, we design shared rollout prefixes and curriculum rollout training to facilitate more effective learning under LQ-anchored flip-mask prediction framework, and the ablation studies in our paper confirm the effectiveness of each individual component.

Table R7. Comparison of LQ conditioning strategies on RealSR with all other settings held fixed.

LQ conditioningPSNR ↑SSIM ↑LPIPS ↓DISTS ↓MANIQA ↑CLIPIQA ↑MUSIQ ↑TOPIQ ↑
w/o current-scale LQ input (Prediction Style)23.070.62400.40170.26280.63890.756170.560.7029
LQ-anchored bit and margin conditioning (GPABC)23.540.63960.40690.26880.66950.783170.660.7266

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.