acceptodds
Under review as a conference paper at ICLR 2027

GB-LSR: Local Spectral Decoding with a Learned Global Bandwidth for Arbitrary-Scale Super-Resolution

Abstract

We present GB-LSR (Global-Bandwidth Local Spectral Representation), a fixed-grid local spectral representation for continuous image decoding. The image domain is partitioned into non-overlapping square patches. Each patch carries coefficients for a truncated Fourier basis, predicted by a single linear projection from shared convolutional-encoder features, and one trainable scalar bandwidth is shared across every patch and every image. As in earlier local spectral decoders, decoding at a continuous coordinate is a fixed-size basis contraction whose cost is set by the spectral cutoff; GB-LSR learns the bandwidth of that basis instead of fixing it. We evaluate an arbitrary-scale super-resolution extension, GB-LSR-Scalar-ASR, against the authors' released LIIF, LTE, and SRNO checkpoints on the same RDN encoder, with every method scored under one protocol and timed in one session per scale, each on one GPU. It runs 1.25x faster than LIIF-RDN at x4 and as fast as SRNO-RDN, whose released code uses 15 times as much peak memory on Urban100. It trails the three encoder-matched baselines by 0.07 to 0.79 dB PSNR-Y in distribution, SRNO-RDN by 0.35 dB on average. Removing the local ensemble raises the speedup to 2.41x over LIIF-RDN and 1.95x over SRNO-RDN at x4, and to 3.00x and 2.41x at x8, without changing PSNR-Y beyond seed variation, at the cost of value jumps at cell boundaries of 0.22 gray levels (of 255) on average at x4. Against the EDSR-baseline checkpoints of five recent methods at x4, GB-LSR-Scalar-ASR scores above or within 0.17 dB on PSNR-Y of LMF, SRNO-EDSR, and OPE-SR-EDSR (1.39 to 6.43 million parameters against 22.02) and 0.14 to 0.57 dB below GSASR and Thera (20.44 and 5.85 million), and has a higher mean LPIPS at x4 than every baseline.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
arbitrary-scale super-resolution, implicit neural representation, local spectral representation, Fourier basis, learned bandwidth, efficient inference

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_6ojyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GB-LSR, a local spectral decoding framework for arbitrary-scale image super-resolution. The method represents local image content using Fourier basis functions and predicts the corresponding coefficients from convolutional features, with a learnable global bandwidth parameter controlling the frequency range. The authors further investigate a variant that removes local ensemble operations to improve inference efficiency. Experiments evaluate reconstruction quality, computational efficiency, bandwidth sensitivity, and boundary continuity across multiple benchmarks.

Critical strengths

1.The proposed framework offers a relatively simple approach to continuous image representation without relying on expensive per-query MLP evaluations. The analysis of local ensemble removal also provides useful insights into the trade-off between inference efficiency and reconstruction continuity.

2.The authors investigate several important design choices, including bandwidth initialization, spectral representation capacity, and local ensemble operations. The inclusion of negative results and boundary discontinuity measurements is particularly valuable, as it helps clarify the practical limitations of the proposed framework.

Critical weaknesses
  1. The central methodological contribution is insufficiently differentiated from existing spectral decoding approaches.

The proposed framework builds upon established ideas in local spectral representations, particularly the use of predicted Fourier coefficients for continuous image reconstruction. Its main distinguishing feature appears to be the introduction of a learnable global bandwidth parameter. However, the experimental evidence does not convincingly demonstrate that learning this parameter provides a meaningful advantage over appropriately selected fixed bandwidths.

In fact, the paper's own ablations indicate that fixed bandwidth configurations can achieve nearly identical reconstruction quality. Furthermore, the learnable bandwidth exhibits sensitivity to initialization, raising additional questions about its practical robustness and optimization benefits. Importantly, the arbitrary-scale super-resolution experiments do not include a sufficiently controlled comparison between learned and fixed bandwidths under otherwise identical conditions.

Consequently, it remains unclear whether the proposed learnable bandwidth constitutes a substantive methodological advance or primarily replaces a manually selected hyperparameter with an additional trainable scalar. This is a fundamental concern because the learnable bandwidth is presented as one of the paper's central contributions.

  1. The claimed efficiency advantages are not convincingly attributable to the proposed methodological innovations.

A substantial portion of the reported inference acceleration comes from removing local ensemble operations, rather than from a demonstrated advantage of the learnable bandwidth mechanism itself. Although the resulting efficiency improvements are practically relevant, the paper does not establish whether similar gains could be achieved by applying the same modification to existing local spectral decoding methods.

Moreover, comparisons with closely related approaches involve differences in encoder architectures, representation capacities, and training configurations, making it difficult to isolate the contribution of the proposed decoder. The default configuration also does not demonstrate a particularly compelling quality-efficiency advantage over strong existing alternatives.

As a result, the paper provides useful empirical observations about architectural simplification, but the evidence remains insufficient to establish that the proposed framework offers a distinct and meaningful improvement over existing spectral representations.

Questions for authors
  1. What is the concrete advantage of learning the global bandwidth rather than using an appropriately selected fixed bandwidth? The current ablations suggest that fixed bandwidth configurations can achieve nearly identical reconstruction quality, while the learned parameter remains sensitive to initialization. Could the authors provide a strictly controlled comparison between learned and fixed bandwidths in the arbitrary-scale super-resolution setting, and clarify what meaningful benefit the learnable formulation provides in terms of reconstruction quality, robustness, or optimization?

  2. Can the authors demonstrate that the reported efficiency improvements arise from the proposed framework rather than from removing local ensemble operations? In particular, could the authors compare against an existing local spectral decoding approach under matched encoder architectures, representation capacities, and training conditions, with both methods trained without local ensemble operations? Such a comparison would be important for establishing whether the proposed method provides an advantage beyond architectural simplification.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 6ojy

We thank the reviewer. We ran both requested controls after submission.

Q2. Matched comparisons without the local ensemble. (a) Matched OPE-SR/GB-LSR comparison (Appendix A.4.2). All four configurations use a 64-channel RDN encoder, a 3x3 coefficient head and OPE-SR's seven normalized modes per axis. At bandwidth 0.5, the decoder is OPE-SR's, under our training setup and initialization. Released OPE-SR-RDN has 147 encoder channels; submitted GB-LSR has 16 modes per axis and a 1x1 head. Normalized renderers match on identical coefficient maps at 0.5; GB-LSR's unnormalized basis needs diagonal coefficient rescaling. We crossed fixed 0.5 versus learning from 0.5 with ensemble on/off in training and evaluation, matching initial weights and data streams within each of three seeds. Mean PSNR-Y equally weights Set5, Set14, B100, Urban100 and DIV2K validation (all 100 images) at x2/x3/x4. Intervals are 95% paired t intervals over seeds:

  • Ensemble removal: +0.023 dB (-0.004 to +0.049) for fixed bandwidth and +0.021 dB (-0.0003 to +0.0413) for learned bandwidth; both are inconclusive.
  • Learned minus fixed: +0.002 dB (-0.0167 to +0.0204) with the ensemble, inconclusive; approximately 0.000 dB (-0.008 to +0.008) without it, equivalent within ±0.02 dB. The interaction, defined as the latter effect minus the former, is -0.002 dB (-0.013 to +0.008), also equivalent.

At x1, an identity diagnostic, ensemble removal lowers mean PSNR-Y by 1.0 dB at fixed 0.5 and 2.8 dB when learning from 0.5. In a separate H200 x4 timing run using the separable raster renderer, geometric-mean latency ratios over Set14/B100/Urban100 are 0.953 end to end and 0.522 for the decoder alone when removing the ensemble at fixed 0.5. Learned/fixed end-to-end latency with the ensemble is 1.001.

(b) LTE ensemble control (Appendix A.4.3). We retrained LTE-RDN with and without the ensemble using its official code, matched initial weights and data streams, and three paired seeds. No-ensemble minus ensemble is -0.005 dB (-0.023 to +0.013) at x2/x3/x4, inconclusive. It is -0.016 dB at x1.25/x1.5/x1.75/x2.5/x3.5 (-0.028 to -0.003) and at x6/x8 (-0.031 to -0.002), averaging the same five datasets. At x1 both models nearly reconstruct the input (an identity diagnostic).

Without the ensemble, LTE has boundary jumps in every tested dataset and seed; at x4, the mean step-corrected jump J(1e-3) is 0.4476 versus 0.0348 gray levels on a 0-255 scale (Table 15). Ensemble removal also causes jumps in (a). In a separate H200 x4 timing run, ensemble removal multiplies LTE's end-to-end latency by 0.4615 (seed range 0.4614-0.4617; geometric mean as in (a)), versus 0.953 for the OPE-style raster decoder in (a). These effects depend on the implementations. In this run, GB-LSR's raster renderer with the ensemble is 1.59x faster than LTE without it. The submitted model's additional speedup came largely from ensemble removal; bandwidth learning does not accelerate inference.

Speed attribution and quality trade-off. In separate H200 x4 timings, the separable raster renderer keeps the ensemble and is 3.59x, 3.48x and 2.91x faster end to end than released LIIF-, LTE- and SRNO-RDN (same geometric mean; Appendix A.4.4). These are dense-raster timings at batch 1. On the same released OPE-SR-RDN weights, it gives a 2.14x speedup over OPE-SR's renderer. This is an implementation improvement for both methods. In (a), fixed-0.5 OPE-style and submitted GB-LSR models, both with the ensemble, have nearly identical end-to-end latency (OPE-style/GB-LSR ratio 0.9998). GB-LSR still has lower in-distribution PSNR-Y and worse mean LPIPS than LIIF-, LTE- and SRNO-RDN (Section 5.1; Figure 2).

Q1. Controlled learned-versus-fixed ASR. With three paired seeds and only trainability changed at each start, learned minus fixed for mean PSNR-Y over those five datasets at x2/x3/x4 is +0.007 dB (95% paired t CI -0.021 to +0.034) from 1.0 and -0.002 dB (-0.029 to +0.025) from 0.5; neither benefit nor equivalence is established.

After these studies, we specified a five-value search before the new runs, including the observed mean learned endpoint 0.882. It selected 0.882 on DIV2K 0801-0850. On fresh paired seeds 3-5, reporting on DIV2K 0851-0900 plus Set5, Set14, B100 and Urban100, fixed minus learned-from-1.0 is -0.003 dB (-0.013 to +0.008) at x2/x3/x4, equivalent within ±0.02 dB for this mean. Both DIV2K halves were previously evaluated. At the pre-specified x1.125/x1.375/x2.25/x2.75/x3.25/x3.75, the five-dataset mean is -0.015 dB (-0.076 to +0.045), inconclusive. Learning remains initialization-sensitive. The revision describes learning as optional joint calibration, without claiming accuracy, robustness or optimization superiority over the selected fixed value (Section 5.3; Appendix A.4.1). The contribution is joint learning of one bandwidth shared across patches and images, plus the native reconstruction study; the renderer is reported separately.

Thank you for the additional experiments. However, I do not think the new results resolve my main concerns. Instead, they further highlight the limitations of the proposed contribution.

  1. Limited value of learnable bandwidth. Neither in the ASR setting nor in the matched OPE-SR setting does the learned bandwidth demonstrate a clear advantage over a properly selected fixed value. The authors also acknowledge that bandwidth learning is merely an optional calibration mechanism and remains sensitive to initialization. Therefore, the central methodological contribution is still insufficiently justified.

  2. Efficiency gains are not convincingly attributable to the proposed method. The authors acknowledge that the original speedup largely comes from removing the local ensemble. Furthermore, the separable raster renderer is essentially an implementation-level optimization that can also be applied to OPE-SR. The matched experiments show nearly identical inference latency between OPE-style and GB-LSR decoders when the local ensemble is retained. Thus, the reported efficiency improvements do not establish a distinctive advantage of the proposed spectral representation itself.

Overall, the additional experiments mainly validate existing design choices and implementation optimizations, rather than demonstrating a substantive benefit from learning the global bandwidth. The fundamental questions regarding methodological novelty and the distinct contribution beyond existing local spectral representations remain unresolved. Therefore, I prefer to maintain my original rating of 2 (Weak Reject).

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

This paper proposes GB-LSR, a local Fourier-based decoder for arbitrary-scale image super-resolution. Unlike implicit neural decoders that use an MLP to predict RGB values at each queried coordinate, GB-LSR predicts local Fourier coefficients and reconstructs pixels through direct evaluation of fixed basis functions. A learnable global bandwidth controls the frequency scale of these basis functions. Experiments evaluate reconstruction quality, inference speed, memory consumption, and different bandwidth configurations.

Critical strengths

1.The Fourier-based decoder replaces per-coordinate neural network evaluation with direct computation using local Fourier coefficients and basis functions. 2.The paper reports inference latency, computational cost, and peak memory, providing a more comprehensive assessment than reconstruction metrics alone.

Critical weaknesses

1.Experiments show that a fixed bandwidth selected appropriately can achieve nearly the same performance as the learnable version. Therefore, the independent benefit of learning the global bandwidth remains unclear. 2.Although the paper includes several relevant approaches, some recent methods are not evaluated. Comparisons across different encoders and model sizes also make it difficult to isolate the contribution of the proposed decoder. 3.GB-LSR underperforms LIIF, LTE, and SRNO when using the RDN encoder in the reported comparisons.

Questions for authors

1.Since arbitrary-scale reconstruction is already supported by existing continuous-coordinate methods, what is the main methodological advance of GB-LSR beyond replacing an MLP decoder with a Fourier-based decoder? 2.How should GB-LSR be positioned relative to existing methods that achieve better PSNR or LPIPS? Can the authors provide a more detailed analysis of the trade-off between inference speed, reconstruction quality, and perceptual quality? 3.Given that a properly selected fixed bandwidth achieves nearly the same performance as the learnable version, what practical or methodological advantage does bandwidth learning provide? Can the authors compare both approaches under a systematic fixed-bandwidth search? 4.How does GB-LSR perform at scales outside the training range compared with recent methods, particularly in terms of both reconstruction quality and inference efficiency?

2 comments
AuthorsOpenReview ↗

Response to Reviewer dMQF

We thank the reviewer. The new results below supplement the submitted comparisons.

Q3 and W1. Learned versus validation-selected fixed bandwidth. After earlier bandwidth studies, we specified the selection rule before the new search runs; 0.882 was already the observed mean learned endpoint. The fixed values were 0.5, 0.7, 0.882, 1.0 and 1.25, with three seeds each (nine new runs, six reused controls). The rule selected 0.882 on DIV2K 0801-0850. Six fresh runs compared it with learning from 1.0 on paired seeds 3-5, reporting on DIV2K 0851-0900 plus Set5, Set14, B100 and Urban100. Both DIV2K halves had been evaluated previously. The equally weighted mean x2/x3/x4 PSNR-Y difference across the five reporting datasets, fixed minus learned, is -0.003 dB (95% paired t CI -0.013 to +0.008), equivalent within ±0.02 dB. Averaging equally over the same datasets at the pre-specified scales x1.125, x1.375, x2.25, x2.75, x3.25 and x3.75 gives -0.015 dB (-0.076 to +0.045), inconclusive. Equivalence applies to the primary mean. Learning remains sensitive to initialization; the revised PDF presents it as optional joint calibration, without claiming superiority over a well-chosen fixed value (Section 5.3, Table 3; Appendix A.4.1).

Q1. Methodological advance. Local Fourier contraction without a per-query network is shared with OPE-SR. GB-LSR adds one trainable global bandwidth and a native reconstruction study of patch size, capacity and per-patch bandwidth. A separate post-submission contribution is separable raster evaluation: on released OPE-SR-RDN weights it matches the official outputs to within 8.4e-7 (maximum absolute difference in normalized RGB, 22 tests) and gives end-to-end speedups of 2.14x at x4 and 3.14x at x8 in separate timing measurements.

To isolate the effect of bandwidth learning, we used a 64-channel RDN encoder, a 3x3 coefficient head and OPE-SR's seven normalized modes per axis, comparing fixed 0.5 (the OPE-SR basis) with learning from 0.5 (Appendix A.4.2). Initial weights and data were matched within each of three seeds. With the ensemble disabled during training and evaluation, learned minus fixed mean PSNR-Y over the five datasets at x2/x3/x4, using all 100 DIV2K validation images, is approximately 0.000 dB (95% paired t CI -0.008 to +0.008), equivalent within ±0.02 dB. This matched-capacity comparison is separate from the submitted model (16 modes per axis) and released OPE training.

W2, W3 and Q2. Quality-efficiency positioning. Averaging equally over five datasets at x2/x3/x4, using all 100 DIV2K validation images, GB-LSR's PSNR-Y minus released LIIF-RDN, LTE-RDN and OPE-SR-RDN is -0.159, -0.236 and +0.090 dB. The submitted-protocol SRNO-RDN comparison is -0.349 dB and +0.0155 LPIPS. These SRNO results cover integer scales only. Thera-RDN Pro (no self-ensemble), LMF-RDN and GSASR-RDN exceed GB-LSR by 0.428, 0.157 and 0.417 dB. GB-LSR has worse mean LPIPS than all seven. These compare three GB-LSR models from the submission with one released checkpoint per baseline. They share the RDN encoder family, but OPE uses 147 rather than 64 features, and released training/normalization choices differ; these are not isolated decoder ablations.

For dense raster rendering at batch size 1 on one H200 at x4, GB-LSR with separable raster evaluation and the local ensemble is faster end to end than the seven released implementations: LIIF 3.59x, LTE 3.48x, SRNO 2.91x, OPE 3.21x, Thera 3.04x, LMF 1.52x and GSASR 8.74x (geometric mean over Set14/B100/Urban100). Thera uses synchronized wall-clock timing in JAX; the other methods use CUDA events in PyTorch. With our renderer, released OPE takes 1.50x GB-LSR's end-to-end time, with its wider encoder. The speed claim is workload-specific: for 100,000 scattered queries on 64x64 inputs at nominal x4, LMF-RDN is slightly faster (latency ratio to GB-LSR 0.97). Figure 2 plots PSNR-Y and LPIPS against latency, distinguishing released implementations from the common GB-LSR renderer.

Q4. Outside the [1,4] training range. At x6/x8, GB-LSR's mean PSNR-Y deficits to LIIF-, LTE-, Thera-, LMF- and GSASR-RDN are 0.116, 0.186, 0.343, 0.114 and 0.366 dB. Their x8 end-to-end latency ratios to GB-LSR are 5.59, 5.43, 4.12, 2.55 and 6.81 in the separate x8 timing measurements.

Across x12/x18/x24/x30 on all 100 DIV2K validation images, GB-LSR minus LIIF is -0.007 dB (95% CI -0.023 to +0.008) and minus LTE -0.065 dB (-0.081 to -0.050). These paired image-bootstrap intervals condition on the trained checkpoints and exclude training variability; the LIIF interval establishes neither a difference nor equivalence. Mean LPIPS remains worse by 0.0361 and 0.0160. A different, fixed-64x64-input workload at x12 gives LIIF/GB-LSR and LTE/GB-LSR latency ratios of 12.44 and 12.10; these are not timings of the DIV2K quality evaluation. We make no domain-generalization claim.

Thank you for conducting the additional experiments. However, I still consider these new results unsatisfactory in addressing my primary concerns, and the novelty of the proposed method remains relatively limited. Therefore, I am inclined to maintain my initial score.

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

This paper proposes GB-LSR, a local spectral representation for arbitrary-scale super-resolution using truncated Fourier bases and a single globally learned bandwidth. Each patch predicts spectral coefficients through a linear projection, enabling continuous-coordinate decoding without a learned per-query network. Experiments compare reconstruction quality and inference efficiency against established SR methods, while additional studies investigate bandwidth learning, spectral cutoff, patch size, and boundary continuity. The method achieves moderate speedups over LIIF and LTE with the same RDN encoder, but generally lower reconstruction accuracy and perceptual quality.

Critical strengths
  1. The fixed spectral decoder offers predictable per-query complexity. Removing the local ensemble achieves a 2.41× speedup over LIIF-RDN at ×4 with negligible average PSNR change.

  2. The paper provides careful experimental documentation, including unified evaluation protocols, multiple training seeds, bandwidth sensitivity analysis, and quantitative measurements of boundary discontinuities.

Critical weaknesses
  1. Limited benefit of learning the bandwidth. The main distinction from existing methods such as OPE-SR is the learnable global bandwidth. However, Appendix A.1.3 shows that fixing the bandwidth at its learned value achieves nearly identical results. In arbitrary-scale SR, the bandwidth also changes little during training and remains sensitive to initialization. It is therefore unclear whether learning the bandwidth provides a meaningful advantage over using a properly selected fixed value.

  2. Limited quality–efficiency advantage. Although GB-LSR improves inference efficiency over some baselines, it achieves lower PSNR than LIIF, LTE, and SRNO with the same RDN encoder, and generally shows weaker perceptual quality. At ×4, Thera achieves higher PSNR, lower LPIPS, fewer parameters, and faster reported inference than the base GB-LSR variant, although the architectures and implementations differ. Removing the local ensemble further improves speed but introduces discontinuities without meaningful accuracy gains. These results leave the overall quality–efficiency trade-off less convincing compared with existing methods.

Questions for authors
  1. Appendix A.1.3 shows that a fixed bandwidth can match the learned bandwidth in native reconstruction. Does this also hold for arbitrary-scale SR? Could the authors provide a comparison using a validation-selected fixed bandwidth under the same training setup?

  2. Appendix A.3 mentions that RDN checkpoints are available for methods such as Thera and LMF. Could the authors evaluate these checkpoints under the same quality and latency protocols to provide a more informative comparison of the accuracy–efficiency trade-off?

2 comments
AuthorsOpenReview ↗

Response to Reviewer tDNZ

We thank the reviewer for the precise questions. Both requested experiments are complete.

Q1. A validation-selected fixed bandwidth under the same ASR setup. We specified the selection rule before the new search runs, after earlier bandwidth studies had found a mean final learned bandwidth of 0.882. We compared five fixed values under the submitted ASR architecture and recipe, three seeds each (nine new runs; six reused controls). On DIV2K 0801-0850, equally averaging PSNR-Y over x1.25, x1.5, x1.75, x2, x2.5, x3, x3.5 and x4, the scores for 0.5, 0.7, 0.882, 1.0 and 1.25 were 36.3482, 36.3535, 36.3583, 36.1634 and 33.9355 dB. The rule selected s* = 0.882; the three reused learned-from-1.0 controls scored 36.3593 dB.

Six additional runs compared fixed s* with bandwidth learned from 1.0 on paired seeds 3-5, evaluated on DIV2K 0851-0900 plus Set5, Set14, B100 and Urban100. Both DIV2K halves had been evaluated previously; this comparison uses new training seeds and new fractional scales on previously evaluated images. The following 95% paired t intervals are over the three seed differences.

  • Primary, the equally weighted mean over five datasets at x2/x3/x4: fixed minus learned is -0.003 dB (-0.013 to +0.008), with the whole interval inside the pre-specified ±0.02 dB margin.
  • Secondary, the mean over five datasets at x1.125, x1.375, x2.25, x2.75, x3.25 and x3.75: -0.015 dB (-0.076 to +0.045), inconclusive. At x1.125, fixed is lower on all five datasets by 0.043-0.134 dB in the seed means. At x1, an identity diagnostic excluded from both outcomes, it is lower on four of five.

The selected fixed value is equivalent to learning from 1.0 for the primary mean; this does not establish equivalence at every scale or on LPIPS. Some fixed choices perform poorly: with seeds 0-2, equally averaging over the five datasets and using all 100 DIV2K validation images, fixed 1.0 is 0.270 dB below learned-from-1.0 at x1.25/x1.5/x1.75/x2.5/x3.5, and fixed 1.25 is 1.150 dB below it at x2/x3/x4. Learning remains initialization-sensitive. Section 5.3 of the revised PDF describes learning as optional joint calibration of the shared bandwidth, without claiming superiority over a well-chosen fixed value (protocol in Appendix A.4.1).

Q2. Thera-, LMF- and GSASR-RDN under the same protocols. Under our quality protocol, we evaluated their released checkpoints against the three GB-LSR models reported in the submission. At x2/x3/x4 (five datasets, using all 100 DIV2K validation images), their mean PSNR-Y advantages over GB-LSR-Scalar-ASR (base) are 0.428 dB (Thera-RDN Pro, no self-ensemble), 0.157 dB (LMF-RDN) and 0.417 dB (GSASR-RDN); their mean LPIPS is lower by 0.0120, 0.0103 and 0.0180. These are released-model comparisons, with training and implementation differences (Appendix A.4.5).

The H200 measurements at x4 and x8 were conducted separately, using the paper's end-to-end protocol and geometric mean of per-dataset latency ratios over Set14, B100 and Urban100. With GB-LSR's ensemble kept, the speedups over those released implementations are respectively 3.04x, 1.52x and 8.74x at x4, and 4.12x, 2.55x and 6.81x at x8. Thera uses synchronized wall-clock timing in JAX; the other methods use CUDA events in PyTorch. These dense-raster results show lower latency at the stated PSNR-Y and LPIPS cost. Figure 2 in the revised PDF distinguishes each method's own released implementation from the shared separable raster implementation applied to GB-LSR and OPE-SR weights. GB-LSR remains below LIIF-, LTE- and SRNO-RDN on in-distribution PSNR-Y. The Thera-EDSR checkpoint of the submission also has fewer parameters, higher PSNR-Y on all five test sets and lower mean LPIPS at x4 (Table 1), and was faster than GB-LSR with its per-query renderer (Table 2). When GB-LSR uses the separable raster renderer, Thera-EDSR takes 2.05x GB-LSR's end-to-end time at x4 (Table 16).

Ensemble removal. The speedups use a post-submission, algebraically equivalent separable raster evaluation of the same trained models (Appendix A.4.4). We no longer use ensemble removal as our main efficiency claim: in separate x4 measurements with the separable raster implementation, removing the ensemble from GB-LSR reduces end-to-end latency by 4.5% (ratio 0.955). It remains an ablation with boundary discontinuities; the submitted variant without the ensemble had a mean x4 jump of 0.22 gray levels out of 255.

Reviewer_tDNZOpenReview ↗

Lack of distinction from existing methods

I appreciate the author providing additional experimental data, which addresses some of my questions regarding the experiments in the paper. However, I still find the paper’s core contributions insufficient, and the distinction from existing methods remains not significant enough; therefore, I am inclined to maintain my original score.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revised PDF: Summary of changes in the revised PDF

Revised PDF. We uploaded a revised PDF; the main text remains within 10 pages. The new evidence is in Section 5.3 and Appendix A.4, and submitted table entries are unchanged. The changes are summarized below.

  1. Learned versus fixed bandwidth (all reviewers). Section 5.3 and Table 3 report the pre-specified fixed-bandwidth search and its confirmation on new seeds; Appendix A.4.1 gives the protocol. The introduction and conclusion now state that learning from 1.0 matches the selected fixed value rather than improving on it, within ±0.02 dB for the primary five-dataset ×2/×3/×4 mean PSNR-Y; the secondary fractional-scale outcome is inconclusive. The limitations continue to make no claim that learning improves on a well-chosen fixed bandwidth. Section 3.4 replaces “stays near its start” with the measured bandwidth endpoints.
  2. Speed attribution (all reviewers). The introduction and related work clarify that local Fourier contraction without a per-query network is shared with OPE-SR. Section 5.3 and Appendix A.4.4 add the algebraically equivalent separable raster renderer, keeping the local ensemble, its ×4 and ×8 timings (Table 16), and its application to released OPE-SR weights. Appendix A.4.4 also adds fixed-input ×12 and scattered-query workload checks. The speed claim is limited to dense raster rendering at batch 1; LMF-RDN is slightly faster on 100,000 scattered queries. Figure 2 plots PSNR-Y and LPIPS against latency, distinguishing released implementations from the common renderer. Ensemble removal remains an ablation.
  3. Matched controls (6ojy, dMQF). Appendix A.4.2 (Table 14) reports a matched factorial with OPE-SR's basis, crossing fixed and learned bandwidth with and without the ensemble; Appendix A.4.3 (Table 15) reports LTE-RDN retrained with and without its ensemble.
  4. Added released baselines (tDNZ, dMQF). Appendix A.4.5 adds results for the released RDN checkpoints of Thera, LMF and GSASR, exact-grid comparisons with released LIIF-, LTE- and OPE-SR-RDN, and paired image-bootstrap intervals for the extreme-scale ladder.
  5. Exploratory extension. Appendix A.4.6 describes the scale-dependent envelope and a combined variant built on it, with their controls (Table 17) and timing (Table 18), and summarizes, regardless of sign, two pre-specified selection criteria that were not met and descriptive error diagnostics.
  6. Scope, interpretation and layout. The setup and reproducibility statements distinguish released-model comparisons, retrained controls and timing sessions; compute totals cover the training runs outside Section 5.3 and Appendix A.4. Appendix A.1.1 now describes the matched-latent comparisons without an interval-based equivalence claim. Appendix A.1.3 removes the submitted matching rule and the claim that training the bandwidth matters only through its final value, reporting the fixed-versus-learned differences descriptively. The abstract and Section 5.2 no longer repeat the ×8 speed and PSNR-Y comparisons, and Section 5.2 omits the ensemble ablation's largest per-cell PSNR-Y change (-0.008 dB on Set14 ×4); Tables 7 to 9 still report these results. The qualitative figure moved to Appendix A.2 (now Figure 7).
AuthorsOpenReview ↗

General response to all reviewers

We thank the reviewers. The completed controls address bandwidth learning and the attribution of speed; the individual replies give the detailed comparisons.

Bandwidth. After earlier bandwidth studies, we specified the fixed-bandwidth selection rule before running the new search. The search over five values, with three seeds each, selected 0.882 on DIV2K 0801-0850; that value was already the observed mean learned endpoint. On fresh paired seeds 3-5, reporting on DIV2K 0851-0900 plus Set5, Set14, B100 and Urban100, fixed minus learned-from-1.0 is -0.003 dB in the equally weighted mean PSNR-Y over the five reporting datasets at x2/x3/x4 (95% paired t CI -0.013 to +0.008). The whole interval lies within the pre-specified ±0.02 dB margin. At the six new scales, the mean over those five datasets is -0.015 dB (-0.076 to +0.045), inconclusive. Both DIV2K halves were previously evaluated; the confirmation uses new training seeds and additional scale factors. Learning remains initialization-sensitive and is an optional calibration, without established superiority over a well-chosen fixed value.

Controlled attribution. The controlled study uses a 64-channel RDN encoder, a 3x3 coefficient head and seven normalized modes per axis, with three paired seeds under our recipe and initialization. It is separate from released OPE-SR training and the submitted GB-LSR model with 16 modes per axis. At x2/x3/x4 on all five datasets, including all 100 DIV2K validation images, learned-from-0.5 minus fixed 0.5 without the ensemble is approximately 0.000 dB (95% paired t CI -0.008 to +0.008), equivalent within ±0.02 dB for this mean. On the same datasets and scales, LTE retrained on three paired seeds gives -0.005 dB (-0.023 to +0.013) for no-ensemble minus ensemble, which does not establish equivalence. Ensemble removal introduces boundary jumps. At x4 on an H200, the retrained LTE model without the ensemble took 0.4615 times the end-to-end latency of the model with it (geometric mean of ratios over Set14, B100 and Urban100).

Quality and speed. The submitted model remains below LIIF-, LTE- and SRNO-RDN on in-distribution PSNR-Y. Released Thera-RDN Pro, LMF-RDN and GSASR-RDN also have higher mean PSNR-Y and lower mean LPIPS under our protocol. A post-submission, algebraically equivalent separable raster implementation keeps GB-LSR's ensemble. In the post-submission H200 x4 measurements of dense raster rendering at batch size 1, it is 3.59x, 3.48x and 2.91x faster end to end than released LIIF-, LTE- and SRNO-RDN, using geometric means over Set14, B100 and Urban100. The same implementation accelerates released OPE-SR-RDN weights by 2.14x over OPE's released renderer; the fixed-0.5 OPE-style model and submitted GB-LSR model, both with the ensemble, have essentially equal end-to-end latency in a separate matched timing run. Learning the scalar is not a speed mechanism, and ensemble removal is retained as an ablation.

Contribution and scope. GB-LSR's contribution is a jointly trained bandwidth shared across patches and images within the existing local spectral decoder family, together with the native-reconstruction design study. The renderer is a separate post-submission implementation improvement demonstrated for GB-LSR and OPE-SR. The individual replies report the quality-efficiency trade-offs, out-of-range scales and workload limitations; they distinguish released-model comparisons from controlled ablations. The new evidence is in Section 5.3 and Appendix A.4 of the revised PDF; Figure 2 shows the quality-latency trade-offs.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.