acceptodds
Under review as a conference paper at ICLR 2027

Graphene: A Lightweight Transformer with Ising Mean-Field Refinement for Image Super-Resolution

Abstract

This paper proposes *Graphene*, a lightweight image super-resolution architecture with two complementary feature refinement mechanisms: Ising-model-inspired local feature refinement using an *Ising Refinement Block* (IRB) and stage-history-based feature refinement using a *History Attention* (HistAttn) module. The IRB reuses attention output features and pre-softmax attention logits to construct signed pairwise token couplings within each local window, applies unrolled mean-field updates, and refines features through a gated residual correction. These signed couplings allow alignment and opposition between latent spins to modulate local feature refinement. HistAttn attends over the current stage output and earlier representations at the same token position and adaptively blends the current stage output with a history-weighted feature aggregate without adding spatial pairwise attention. For *classical SR* and *lightweight SR* on Urban100 at , Graphene and Graphene-light outperform the state-of-the-art methods IET and IET-light by **0.24 dB** and **0.38 dB**, respectively, with **34.3%/13.7% fewer parameters**, **44.8%/44.4% fewer FLOPs**, **87.9%/84.5% lower latency**, and **78.3%/87.5% lower peak memory**. *Real-world SR* comparisons further show that Graphene recovers fine details and preserves textures across diverse degradations. If the paper is accepted, the code will be released by the camera-ready deadline.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image super-resolution, efficient vision Transformers, Ising model, mean-field inference, feature refinement, history attention

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_5mrFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper proposes Graphene, a lightweight Transformer for image super-resolution with two feature refinement modules: Ising Refinement Block (IRB) and History Attention (HistAttn). IRB uses pre-softmax attention logits to construct signed pairwise interactions and applies mean-field updates to control local residual refinement. HistAttn adaptively aggregates features from different network stages. Experiments show that Graphene achieves competitive reconstruction quality with reduced computational costs.

Critical strengths
  1. IRB derives signed local interactions from attention logits and leverages mean-field updates to gate residual corrections, refining fine local structures coordinately.
  2. HistAttn adaptively reuses representations across stages at each token position, aggregating cross-depth features without additional spatial pairwise attention.
  3. The paper compares different IRB configurations and evaluates adaptive history selection against static weighting and simple skip connections.
Critical weaknesses

1.The motivation for the IRB module is insufficiently explained, and the principles of the method are not clearly stated. 2. Some ablation experiments showed only minor improvements. Compared to using only binary gates, the full IRB model increased PSNR by only 0.04 dB, while HistAttn increased PSNR by only 0.03 dB compared to static historical weights. 3. The argument regarding complementarity is not sufficiently substantiated. Although the combined model achieved the highest PSNR, its performance was only 0.05 dB better than that of IRB used alone. The results indicate that the modules can be used in combination but do not demonstrate strong synergistic effects. 4. The efficiency claims require further substantiation. In the reported component ablation experiments, IRB added approximately 2.4M parameters and 20.4G FLOPs. Furthermore, its reliance on numerical values for pre-softened maximum attention may limit its compatibility with optimized attention implementation schemes.

Questions for authors
  1. Why is Ising mean-field refinement necessary for local feature refinement? Can simpler input-dependent gating or local interaction mechanisms achieve comparable results?
  2. How do the authors distinguish functional complementarity from actual performance synergy? Can they provide stronger evidence that HistAttn consistently improves performance when IRB is already enabled?
  3. Could the author provide a more comprehensive comparison with the latest image super-resolution methods published over the past two years? In particular, a comparison with recent Transformer-based and lightweight super-resolution methods would help illustrate Graphene’s advantages in terms of reconstruction quality, computational efficiency, and methodological innovation.
3 comments
AuthorsOpenReview ↗

No simpler mechanism matches IRB, at lower cost than recent methods

W1 / Q1. Motivation and principle of IRB; can simpler mechanisms compete? As requested, Section 3.2 now opens with a Motivation paragraph\color{blue}{\textbf{Section 3.2 now opens with a Motivation paragraph}}. No simpler mechanism we tested matches IRB.\color{blue}{\textbf{No simpler mechanism we tested matches IRB.}}

  • Motivation. PSA's softmax weights are nonnegative and only mix value features. IRB reuses PSA's logits to coordinate how strongly each token is corrected: centering and symmetrization give signed, reciprocal couplings, so each token's correction depends on its partners.
  • Principle. The couplings and unary fields define an Ising energy (Eq. 5). IRB unrolls its mean-field stationarity condition (Eq. 6, Proposition A.4) for three steps, and the final magnetization sets a bounded gate gi∈[0.269,0.731]g_i\in[0.269,0.731] on a token-wise residual correction (Eq. 7, Propositions A.5–A.6, checked in Lean).
  • Simpler gating, same residual MLP (Table 1(b)). Constant gate 30.85 dB; input-dependent unary gate 30.89 dB; nonnegative and asymmetric couplings at identical cost 30.90 and 30.86 dB; full IRB 30.93 dB.
  • Simpler local interaction at the same compute. A larger window costs the same 20.4G FLOPs as IRB but gains 0.19 dB instead of 0.55 dB\color{blue}{\textbf{A larger window costs the same 20.4G FLOPs as IRB but gains 0.19 dB instead of 0.55 dB}} (Table 1(a)). HAT already has channel attention and overlapping cross-window attention, and IRB still improves it by 0.16–0.94 dB at ×2 (Table 9).

W2. Small increments (IRB vs. unary gate; HistAttn vs. static weights). These numbers compare single design choices at near-identical cost; the module contributions are much larger.\color{blue}{\textbf{the module contributions are much larger.}}

  • Module level. IRB adds 0.55 dB to the backbone (0.40 dB with HistAttn present), HistAttn 0.20 dB (0.05 dB with IRB present), and both 0.60 dB, 7.5× the six-seed SD\color{blue}{\textbf{0.60 dB, 7.5× the six-seed SD}} (Table 1(a), Table 8).
  • Choice level. The couplings cost 1.0G FLOPs over the unary gate (0.76% of the model); static weights and HistAttn differ by 0.2G FLOPs (Tables 1(b), 2).
  • Benchmark level. IET improved on ATD and PFT by 0.21 and 0.23 dB on Urban100 ×4; Graphene improves on IET by 0.24 dB (Table 3).

W3 / Q2. Complementarity versus synergy. We claim complementarity, not super-additive synergy\color{blue}{\textbf{We claim complementarity, not super-additive synergy}}. IRB acts within a window at one depth; HistAttn acts across depth at one token position. Each helps when the other is present:

  • HistAttn with IRB enabled: +0.05 dB (30.88 → 30.93 dB; Table 1(a)), and the highest of four history configurations, all with IRB: no fusion 30.88, shallow skip 30.89, static weights 30.90, HistAttn 30.93 dB (Table 2).
  • IRB with HistAttn enabled: +0.40 dB (30.53 → 30.93 dB; Table 1(a)).

W4. IRB cost and compatibility with optimized attention. As requested, every method is now measured in two settings\color{blue}{\textbf{every method is now measured in two settings}} (BF16 compiled: Section 4.1, Figure 1, Tables 3–4\color{blue}{\textbf{BF16 compiled: Section 4.1, Figure 1, Tables 3–4}}; FP32 eager: Appendix E.3, Figure 8, Tables 15–16\color{blue}{\textbf{FP32 eager: Appendix E.3, Figure 8, Tables 15–16}}). In both, Graphene stays on the accuracy–cost Pareto front at every input size.\color{blue}{\textbf{In both, Graphene stays on the accuracy–cost Pareto front at every input size.}}

  • 2.4M parameters and 20.4G FLOPs are the whole IRB, all 36 blocks with residual MLPs; the couplings alone cost 1.0G FLOPs\color{blue}{\textbf{the couplings alone cost 1.0G FLOPs}} (Table 1). With IRB, Graphene still has fewer parameters and FLOPs than every compared method published since 2024\color{blue}{\textbf{fewer parameters and FLOPs than every compared method published since 2024}}, at every scale (Tables 3, 13).
  • Larger inputs. At 320×180 and 1280×720, Graphene is 6.3–7.3× faster than IET in BF16 (Table 4) and 2.9–3.0× in FP32 (Table 16), and every faster model is at least 0.45 dB lower.
  • Optimized attention. The Limitations (Section 4.5) state that IRB needs the pre-softmax logits, which limits FlashAttention-style kernels; however, as we already stated, its computational cost is lower than state-of-the-art methods\color{blue}{\textbf{its computational cost is lower than state-of-the-art methods}}.

Q3. Comparison with recent methods. Table 3 and Appendix E (full comparison: Tables 13–14) already include the latest models\color{blue}{\textbf{Table 3 and Appendix E (full comparison: Tables 13–14) already include the latest models}}: IPG and ATD (CVPR 2024), MambaIR (ECCV 2024), MambaIRv2 and PFT (CVPR 2025), IET (ECCV 2026) and UCAN (CVPR 2026), and the lightweight versions of ATD, MambaIR, MambaIRv2, PFT and IET.

  • Accuracy–cost comparison. Against IET and IET-light, the strongest PSNR-oriented methods, Graphene and Graphene-light gain 0.24 and 0.38 dB on Urban100 ×4 with\color{blue}{\textbf{Graphene and Graphene-light gain 0.24 and 0.38 dB on Urban100 ×4 with}} 34.3%/13.7% fewer parameters, 44.8%/44.4% fewer FLOPs,\color{blue}{\textbf{34.3%/13.7% fewer parameters, 44.8%/44.4% fewer FLOPs,}} 83.9%/95.9% lower latency and 63.7%/92.0% lower peak memory\color{blue}{\textbf{83.9%/95.9% lower latency and 63.7%/92.0% lower peak memory}} (Table 3).
  • Methodological innovation. Both modules reuse quantities a windowed Transformer already computes, attention logits and stage outputs, instead of adding another pairwise attention (Sections 3.2–3.3).

Thank you for the response. After reviewing the additional experimental results and the comments from other reviewers, the authors address some of my concerns; however, I still have questions regarding the motivation and principles behind this method, and the performance gains resulting from the core operations are limited. Therefore, I retain my original rating.

AuthorsOpenReview ↗

Graphene breaks the rule that higher PSNR must be bought with higher cost

Thank you for your review. If a specific point remains unclear, please tell us which one, and we will answer it directly.

Given the experience behind your confidence rating, you will recognize that Graphene moves the accuracy–cost Pareto frontier forward. Graphene breaks the rule that higher PSNR must be bought with higher cost.\color{red}{\textbf{Graphene breaks the rule that higher PSNR must be bought with higher cost.}} If there is anything you would like to know more about, please let us know; we are confident that we can answer even the subtle points.

Reviewer_KhNkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Graphene, a lightweight image super-resolution architecture that improves reconstruction quality through two complementary feature refinement mechanisms. The Ising-model-inspired local refinement block (IRB) enhances local texture representation through signed pairwise interactions, while the History Attention module (HistAttn) integrates features from the current and earlier stages to enable efficient cross-stage information interaction. Experimental results demonstrate that Graphene outperforms existing methods in both classical and lightweight super-resolution tasks while significantly reducing parameter count, computational cost, inference latency, and memory consumption, achieving an effective balance between reconstruction quality and computational efficiency.

Critical strengths
  1. The paper is easy to understand and follow.
  2. The proposed method achieves a good balance between accuracy and computational efficiency.
Critical weaknesses
  1. The full IRB achieves only a 0.04 dB improvement over the Unary-only gate, which does not incorporate Ising-based pairwise token couplings.

  2. Why should feature refinement in image super-resolution be governed by Ising mean-field inference? While the Ising model provides an interesting form of nonlinear iteration, it remains unclear what unique advantages it offers over conventional graph neural network message passing or iterative gating mechanisms.

  3. Even if certain positions are set to zero by the attention mask, they may become nonzero again after row-centering. This means that originally masked token pairs may reintroduce interactions in the Ising coupling graph.

  4. The visualization results show that HistAttn almost exclusively selects shallow features. Does HistAttn genuinely learn effective dynamic history selection, or does it primarily function as a sophisticated shallow-feature skip connection?

Questions for authors
  1. Does the marginal 0.04 dB improvement justify the additional complexity of Ising-based pairwise couplings?

  2. What unique advantages does Ising mean-field inference offer over conventional message passing or iterative gating mechanisms?

  3. Could row-centering reintroduce interactions between masked token pairs? How is the masking behavior preserved?

  4. Does HistAttn learn meaningful dynamic history selection, or does it mainly act as a sophisticated shallow-feature skip connection?

3 comments
AuthorsOpenReview ↗

The couplings are almost free and matter, the mask holds, and history selection is dynamic

W1 / Q1. Are the pairwise couplings worth the added complexity? Yes. The couplings are almost free, and at identical cost their rule measurably matters.\color{blue}{\textbf{Yes. The couplings are almost free, and at identical cost their rule measurably matters.}}

  • Cost. Over the unary-only gate, the couplings add 1.0G FLOPs (0.76% of the model) and no change in the reported parameter count (12.9M; Table 1(b)). J\mathbf{J} is built from logits PSA has already computed, so no new query–key computation is needed (Section 3.2).
  • The rule matters at the same cost. Nonnegative and asymmetric couplings have the same parameters and FLOPs as full IRB (12.9M, 132.0G) and score 0.03 and 0.07 dB lower (Table 1(b)).
  • Paired evidence. Starting from the same pretrained ×4 model with identical data, intact IRB stays ahead of J=0\mathbf{J}=0 at every checkpoint, with every all-pixel 95% image-level interval above zero, and leads on 67 of 100 crops at the end (Table 10, Appendix C.4).
  • The whole block. IRB adds 0.55 dB at the same window, 2.9× the gain of a larger window for the same 20.4G FLOPs\color{blue}{\textbf{2.9× the gain of a larger window for the same 20.4G FLOPs}} (Table 1(a)), and +0.94 dB on Urban100 ×2 inside HAT (Table 9).

The six-seed SD of 0.08 dB (Table 8) is the unpaired run-to-run variation of one configuration. The Table 1(b) controls are paired (matched shared-parameter initialization and minibatches), so this SD is not the uncertainty of their differences.


W2 / Q2. Why Ising mean-field inference rather than message passing or iterative gating? We do not claim that SR requires Ising physics. The mean-field update gives IRB three properties at a cost the alternatives do not match.\color{blue}{\textbf{The mean-field update gives IRB three properties at a cost the alternatives do not match.}}

  • Signed, reciprocal interactions from logits already computed. Softmax weights are nonnegative, so attention cannot express that two tokens should be corrected in opposite ways. Centering and symmetrizing PSA's logits gives signed, reciprocal couplings with no new affinity computation (Section 3.2).
  • Scalar messages. IRB passes one scalar per token, so one mean-field round costs 2n2n FLOPs per token. Graph neural network message passing with C-channel features\color{blue}{\textbf{Graph neural network message passing with}\ C\textbf{-channel features}} over the same graph costs 2nC: 180× more in the full model\color{blue}{\textbf{over the same graph costs}\ 2nC\textbf{: 180× more in the full model}} (C=180C=180, Table 6). Table 3 includes such a network for SR, IPG. Graphene is 0.54 dB above it, while IPG needs 31% more parameters, 96% more FLOPs, 36% more latency and 105% more peak memory (12% and 88% in FP32 eager inference, Table 15).
  • Interactions are what make iteration useful. With J=0\mathbf{J}=0, every update of Eq. 6 returns tanh⁡(βhi)\tanh(\beta h_i), so iterative gating collapses to the unary-only gate, 0.04 dB below full IRB (Table 1(b)). The update iterates the mean-field self-consistency condition of the Ising energy in Eq. 5 (Proposition A.4), with provably bounded magnetizations and gates (Propositions A.5–A.6), checked in Lean (Appendix A.6).

The learned couplings follow image structure. In one IRB (Stage 6, Block 5), the 1,822 most edge-like and the 1,822 most texture-like of 18,221 Urban100 windows have 6.3× and 7.8× the mean coupling magnitude of the 1,822 flattest (Section 4.4, Appendix D.1).


W3 / Q3. Masked pairs after row-centering. Yes, and the paper states it (Section 3.2, Appendix B.2): after centering, the coupling at a masked pair can be nonzero, but no feature and no query–key affinity crosses the mask.\color{blue}{\textbf{but no feature and no query–key affinity crosses the mask.}}

  • No feature crosses the mask. IRB aggregates no features: its correction R(yi)R(y_i) uses token ii alone, and its input yy is the output of the masked PSA. Only the scalar gate gi∈[0.269,0.731]g_i\in[0.269,0.731] can depend on a masked partner (Eq. 7).
  • No pairwise affinity crosses the mask. Masked entries are zeroed before centering (Eq. 2), so at a masked pair Eq. 3 gives Jij=−γ2(μi+μj)J_{ij}=-\tfrac{\gamma}{2}(\mu_i+\mu_j): a sum of two per-token row means with no query–key term between ii and jj.
  • Masked pairs are confined to border windows. They occur only in shifted-window blocks, in the windows that wrap around the image border (Appendix B.2). IRBs in non-shifted blocks have none.

W4 / Q4. Dynamic selection or a shallow skip? Dynamic. Near-exclusive shallow selection occurs at only two of the five insertions.\color{blue}{\textbf{Dynamic. Near-exclusive shallow selection occurs at only two of the five insertions.}}

  • The exact skip control loses. A learned shallow skip at the same insertions scores 30.89 dB, static history weights 30.90 dB, HistAttn 30.93 dB, all with IRB enabled (Table 2).
  • Selection changes with depth (Table 11). Entering Stage 2: 59.3% shallow, 40.7% current. Stages 3–4: 96.9–98.8% shallow. Stage 5: 55.5% shallow, 41.6% intermediate, 2.9% current. Stage 6: 17.8% shallow, 41.4% intermediate, 40.8% current. A skip connection cannot express the last two.
  • Selection varies across tokens. The token-level SD is 0.179 for the shallow weight entering Stage 2 and 0.106 for the current weight entering Stage 6 (Appendix D.2).

Thank you for the authors' response. The authors have addressed most of my concerns. However, the 0.04 dB improvement is indeed marginal. Therefore, I will maintain my original score.

AuthorsOpenReview ↗

You have misread the gain: it is 0.55 dB, not 0.04 dB

We sincerely thank you for taking the time to evaluate our manuscript. We truly understand what it means to devote time to reviewing. We are also glad that most of your concerns are addressed. However, we do not believe that an expert in this field would give this score, for the following reasons: We believe that, among the state-of-the-art methods we compare, ours is distinguished by\color{red}{\textbf{We believe that, among the state-of-the-art methods we compare, ours is distinguished by}}

  1. the thoroughness of its theoretical background (Appendix A), extended ablations (Appendices C–D)\color{red}{\textbf{the thoroughness of its theoretical background (Appendix A), extended ablations (Appendices C–D)}} and latency and memory evaluation (Section 4.1, Appendix E.3); and\color{red}{\textbf{and latency and memory evaluation (Section 4.1, Appendix E.3); and}}
  2. Graphene’s break from the accuracy–cost trade-off (table below; full comparison in Table 3),\color{red}{\textbf{Graphene's break from the accuracy–cost trade-off (table below; full comparison in Table 3),}} which actually matters on real hardware.\color{red}{\textbf{which actually matters on real hardware.}}

Until now, state-of-the-art accuracy has been bought with cost. However, Graphene breaks this trend with a larger PSNR gain at a lower cost\color{red}{\textbf{Until now, state-of-the-art accuracy has been bought with cost. However, Graphene breaks this trend with a larger PSNR gain at a lower cost}}, as the example in the table below shows (Urban100 ×4, Table 3; BF16 compiled, 64×64 LR input; parentheses: relative to IPG).

YearMethodParams (M)FLOPs (G)Latency (ms)Memory (MiB)Urban100 ×4 (dB)
2024IPG (reference)17.0268.511.31394.828.13
2024ATD22.0 (+29.4%)\color{red}{\textbf{(+29.4%)}}263.5 (−1.9%)\color{blue}{\textbf{(−1.9%)}}9.00 (−20.4%)\color{blue}{\textbf{(−20.4%)}}123.0 (−68.8%)\color{blue}{\textbf{(−68.8%)}}28.22 (+0.09)\color{blue}{\textbf{(+0.09)}}
2025PFT19.8 (+16.5%)\color{red}{\textbf{(+16.5%)}}188.5 (−29.8%)\color{blue}{\textbf{(−29.8%)}}114.99 (+916.7%)\color{red}{\textbf{(+916.7%)}}533.2 (+35.1%)\color{red}{\textbf{(+35.1%)}}28.20 (+0.07)\color{blue}{\textbf{(+0.07)}}
2026IET19.8 (+16.5%)\color{red}{\textbf{(+16.5%)}}248.1 (−7.6%)\color{blue}{\textbf{(−7.6%)}}51.68 (+356.9%)\color{red}{\textbf{(+356.9%)}}530.4 (+34.3%)\color{red}{\textbf{(+34.3%)}}28.43 (+0.30)\color{blue}{\textbf{(+0.30)}}
Graphene13.0 (−23.5%)\color{blue}{\textbf{(−23.5%)}}137.0 (−49.0%)\color{blue}{\textbf{(−49.0%)}}8.34 (−26.3%)\color{blue}{\textbf{(−26.3%)}}192.6 (−51.2%)\color{blue}{\textbf{(−51.2%)}}28.67 (+0.54)\color{blue}{\textbf{(+0.54)}}

Despite this low cost, Figure 5 visually shows that Graphene alone reproduces the dense repeated lines of the HR crop.\color{red}{\textbf{Despite this low cost, Figure 5 visually shows that Graphene alone reproduces the dense repeated lines of the HR crop.}}


You have misread the gain: IRB’s gain is 0.55 dB (Table 1(a)), not 0.04 dB. The 0.04 dB isolates one term inside IRB\color{red}{\textbf{You have misread the gain: IRB's gain is 0.55 dB (Table 1(a)), not 0.04 dB. The 0.04 dB isolates one term inside IRB}}, the pairwise couplings, against a gate that already keeps IRB's unary field and residual MLP; it tests the mechanism, not the size of our contribution, which is measured by IRB as a whole:

  • In Graphene. At the same added 20.4G FLOPs, IRB raises PSNR by 0.55 dB (30.33 to 30.88 dB), 2.9× the 0.19 dB from enlarging the window from 16 to 32 (Table 1(a)).
  • In another network. Added to HAT, IRB improves all five ×2 PSNR values by 0.16–0.94 dB (Table 9).

Even as a mechanism test, the 0.04 dB is not marginal:

  • It is real, not noise. In a paired run from the same pretrained model with identical data, intact IRB stays ahead of J=0J=0 at every checkpoint, with every all-pixel 95% image-level interval above zero, and intact IRB leads on 67 of 100 crops (Table 10, Appendix C.4).
  • It comes from the Ising rule, not from extra compute. At identical parameters and FLOPs, asymmetric couplings fall below even the unary-only gate (30.86 vs 30.89 dB), and nonnegative couplings recover only 0.01 dB (30.90 dB). Only signed, symmetric couplings, the Ising form, give the 0.04 dB.
  • It is cheap. The couplings cost 1.0G FLOPs (0.76% of the model) and no extra parameters: 0.04 dB for 1.0G FLOPs, against 0.19 dB for 20.4G FLOPs when the window grows from 16 to 32 (Table 1(a)), 4.3× the PSNR per FLOP of enlarging the window.
Reviewer_qL7dOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Graphene augments a PSA-based SR Transformer with an Ising Refinement Block, which converts attention logits into signed couplings and uses three mean-field updates to gate residual features, and History Attention, which combines representations across stages. Classical and lightweight variants report strong reconstruction results with reduced parameters and FLOPs relative to IET.

Critical strengths
  • The paper is clearly written and easy to follow.
  • The reported reconstruction results are strong across multiple datasets and scales. Graphene-light achieves 27.79 dB on Urban100 at ×4 with about 692K parameters, suggesting a promising balance between model size and reconstruction quality.
  • The paper includes extensive experiments and ablation studies that help clarify the contribution of each module.
Critical weaknesses
  • Graphene uses 128×128 LR training patches, whereas the compared DF2K configurations of SRFormer and IET use 64×64. Graphene-light also uses DF2K, while IET-light's official recipe uses DIV2K. Since baseline accuracy is taken from prior reports, the improvements cannot be cleanly attributed to the architecture. I could not find a controlled training-crop ablation.
  • Appendix B.4 benchmarks Graphene with BF16 autocast and torch.compile, but baselines with FP32 eager inference. This mixes architectural efficiency with precision and compiler effects, weakening the headline latency and memory comparisons. Profiling is also limited to 64×64 LR inputs. Although reconstruction metrics use full images, no larger-resolution efficiency measurements establish whether the reported advantage persists, particularly when the dense window couplings and stored history occupy more memory.
Questions for authors
  1. How much of the gain remains with 64×64 training crops and matched datasets and optimization settings? Please include a matched DIV2K comparison for Graphene-light and clarify the provenance of the baseline checkpoints/results.
  2. Can the main latency and memory comparisons be repeated under matched precision and execution modes, at both 64×64 and larger rectangular LR sizes such as 320×180 and 1280×720?
3 comments
AuthorsOpenReview ↗

Matched evidence for the architecture, matched settings for the cost

W1 / Q1. Training recipe and baseline provenance. As you requested, we added the 64×64-crop and optimizer comparisons for Graphene to Appendix E.1\color{blue}{\textbf{we added the 64×64-crop and optimizer comparisons for Graphene to Appendix E.1}} and the DIV2K comparison for Graphene-light to Appendix E.2\color{blue}{\textbf{and the DIV2K comparison for Graphene-light to Appendix E.2}}. Our architecture claims rest on comparisons under one training recipe, and training choices do not change inference cost.

  • Provenance. Baseline PSNR/SSIM are the values published in the original papers, the standard SR protocol (Appendix B.3). Every latency and memory value is our own measurement of the released code under one protocol\color{blue}{\textbf{Every latency and memory value is our own measurement of the released code under one protocol}} (Section 4.1, Appendix E.3).
  • Matched evidence for the architecture. Tables 1–2 train every variant from scratch under one recipe. IRB and HistAttn add 0.60 dB, 7.5× the six-seed SD\color{blue}{\textbf{0.60 dB, 7.5× the six-seed SD}} (Table 1(a), Table 8); the lowest of the six full-model seeds (30.81 dB) is still 0.48 dB above the backbone. After 50,000 training iterations, HAT+IRB beats HAT by 0.16–0.94 dB on all five ×2 benchmarks (Appendix C.3, Table 9).
  • Cost is fixed by the architecture. Crops, data and optimizer do not change inference cost. Against IET, Graphene uses 34.3% fewer parameters and 44.8% fewer FLOPs, with 6.2× lower latency and 2.75× lower peak memory\color{blue}{\textbf{6.2× lower latency and 2.75× lower peak memory}} (Table 3). Against IET-light, Graphene-light uses 13.7% fewer parameters and 44.4% fewer FLOPs, with 24.1× lower latency and 12.4× lower peak memory\color{blue}{\textbf{24.1× lower latency and 12.4× lower peak memory}} (Table 3).
  • Crop size, optimizer and training data. Changing only the LR crop from 64×64 to 128×128 raises Graphene's PSNR by at most 0.08 dB (largest on Urban100), and using Muon rather than Adam alone by at most 0.05 dB. Retrained on DIV2K, Graphene-light loses at most 0.23 dB. Even trained on DIV2K, Graphene-light leads every lightweight baseline by at least 0.15 dB on Urban100 ×4. Our focus is PSNR-oriented classical SR at a lightweight computational cost.

W2 / Q2. Matched precision and larger inputs. As you requested, every latency and peak-memory comparison is now repeated under matched precision and execution\color{blue}{\textbf{every latency and peak-memory comparison is now repeated under matched precision and execution}} (BF16 compiled: Figure 1, Tables 3–4 and 13–14\color{blue}{\textbf{BF16 compiled: Figure 1, Tables 3–4 and 13–14}}; FP32 eager: Figure 8, Tables 15–16\color{blue}{\textbf{FP32 eager: Figure 8, Tables 15–16}}), at 64×64, 320×180 and 1280×720 LR inputs (64×64: Figures 1 and 8, Tables 3 and 13–15\color{blue}{\textbf{64×64: Figures 1 and 8, Tables 3 and 13–15}}; 320×180 and 1280×720: Tables 4 and 16\color{blue}{\textbf{320×180 and 1280×720: Tables 4 and 16}}). In every one of these figures and tables, Graphene and Graphene-light stay on the accuracy–cost Pareto front.\color{blue}{\textbf{In every one of these figures and tables, Graphene and Graphene-light stay on the accuracy–cost Pareto front.}}

  • BF16 compiled inference, every method (Section 4.1). Figure 1, Tables 3–4 and Tables 13–14 now measure every method with one script at batch size 1. At 64×64, latency is the median of seven fresh-process runs of 50 warm-up and 200 timed passes, and peak memory the median of three runs. IET now runs NATTEN's CUDA kernels, as its released code requires, instead of a pure-PyTorch fallback. Peak memory now counts one shared copy of each shifted-window mask, as the released code does (FP32 eager: 415.2 → 278.1 MiB for Graphene, bit-identical outputs).
×4, BF16 compiledLatency (ms)Peak memory (MiB)Urban100 (dB)
IET51.68530.428.43
Graphene8.34 (−83.9%)192.6 (−63.7%)28.67
IET-light38.52258.827.41
Graphene-light1.60 (−95.9%)20.8 (−92.0%)27.79
  • Larger inputs in BF16 (Table 4, Section 4.5; median of three runs of 5 warm-up and 20 timed passes). Graphene is 6.3× and 7.3× faster than IET at 320×180 and 1280×720, with 3.3× and 3.5× lower peak memory\color{blue}{\textbf{Graphene is 6.3× and 7.3× faster than IET at 320×180 and 1280×720, with 3.3× and 3.5× lower peak memory}}, and Graphene-light is 21.6× and 24.3× faster than IET-light with 7.6× lower peak memory\color{blue}{\textbf{Graphene-light is 21.6× and 24.3× faster than IET-light with 7.6× lower peak memory}}.
  • FP32 eager inference, every measurement repeated (Appendix E.3: Table 15 and Figure 8 at 64×64, Table 16 at 320×180 and 1280×720). Against IET, Graphene is 1.24×, 3.0× and 2.9× faster with 2.54×, 3.3× and 3.5× lower peak memory\color{blue}{\textbf{Graphene is 1.24×, 3.0× and 2.9× faster with 2.54×, 3.3× and 3.5× lower peak memory}} at 64×64, 320×180 and 1280×720. Graphene-light is the fastest lightweight model at both larger sizes\color{blue}{\textbf{Graphene-light is the fastest lightweight model at both larger sizes}}.
  • Pareto front in both settings. In BF16 and FP32, at every input size, every model that is faster or uses less memory is at least 0.45 dB lower on Urban100 ×4\color{blue}{\textbf{every model that is faster or uses less memory is at least 0.45 dB lower on Urban100 ×4}} (Tables 3–4, 15–16).
  • Why the advantage persists. Windows have a fixed size, so operations and stored coupling and history entries grow linearly with image area (Section 3.3). The coupling matrix holds nn entries per token, fewer than the 1.5n1.5n attention logits PSA already materializes.
Reviewer_qL7dOpenReview ↗

Update

Thank you for your response. The additional experiments addressed some of my concerns. However, the original paper does contain some unfair comparisons, and some recent methods are also missing from the comparisons, so I incline to maintain the initial score.

AuthorsOpenReview ↗

Unifying the inference setting widens Graphene-light's advantage; the state of the art is already included and every omitted method scores lower

Measurement protocol. We firmly disagree that the comparison was unfair. Unifying the inference setting widens Graphene-light’s advantage\color{red}{\textbf{Unifying the inference setting widens Graphene-light's advantage}} over IET-light, from 6.5× faster with 8.0× less memory to 24.1× and 12.4×.

The submission stated the protocol explicitly (Section 4.1, Appendix B): each baseline ran with the inference setting of its official code (FP32 eager), and Graphene with its training precision (BF16 compiled).

Unifying the protocol changes none of our conclusions (BF16 compiled: Figure 1, Tables 3–4 and 13–14; FP32 eager: Figure 8, Tables 15–16). Against IET, Graphene remains 6.2× faster with 2.75× less memory in BF16 compiled (8.3× and 4.6× in the submission) and 1.24× faster with 2.54× less memory in FP32 eager. In both settings, Graphene and Graphene-light stay on the accuracy–cost Pareto front. The original protocol was disclosed and did not create our advantage.\color{red}{\textbf{The original protocol was disclosed and did not create our advantage.}}


Recent methods. Image super-resolution has several subcategories, but our method and our baselines are high-PSNR-oriented,\color{red}{\textbf{Image super-resolution has several subcategories, but our method and our baselines are high-PSNR-oriented,}} and in this category IET (ECCV 2026) and UCAN (CVPR 2026) are the latest state of the art\color{red}{\textbf{and in this category IET (ECCV 2026) and UCAN (CVPR 2026) are the latest state of the art}}. We are confident that we made every effort to include all recent methods that they compare against.

The few methods that our baselines also compare against but our tables do not list all score lower than ours on Urban100 ×4: DAT (in MambaIRv2's table) reaches 27.87 dB against Graphene's 28.67 dB, and ESC, ASID-D8 and HiT-SRF (in UCAN's table) reach 27.07, 26.89 and 26.80 dB against Graphene-light's 27.79 dB. No omitted method changes any conclusion.\color{red}{\textbf{No omitted method changes any conclusion.}}

We searched for such methods ourselves and also with Fable 5.2 Max, Opus 5.5 Max, GPT-6 Astra Max,\color{red}{\textbf{We searched for such methods ourselves and also with Fable 5.2 Max, Opus 5.5 Max, GPT-6 Astra Max,}} GPT-6.1 Sol Max, GPT-6.1 Terra Max, Gemini 4 Carbon, Gemini 4 Argon and Gemini 4 Hydrogen.\color{red}{\textbf{GPT-6.1 Sol Max, GPT-6.1 Terra Max, Gemini 4 Carbon, Gemini 4 Argon and Gemini 4 Hydrogen.}} Once more, please check which methods our cited high-PSNR-oriented state-of-the-art baselines compare against. For reference, we have no affiliation with the authors of these papers. If you have a specific method in mind, please name it and we will add it.\color{red}{\textbf{If you have a specific method in mind, please name it and we will add it.}}

Comments on the paper

1 comment
AuthorsOpenReview ↗

Beyond the accuracy–cost Pareto front in BF16 and FP32, under one protocol

Acknowledgment. We thank the reviewers and the AC for their valuable time and feedback. We answer every point, state exactly what was revised, and mark key points in blue. Under one protocol for every method, \href\href{https://anonymous.4open.science/r/Graphene-8F2C/}{\color{blue}{\textbf{reproducible from our repository}}}, Graphene moves the accuracy–cost Pareto frontier forward, whether all methods run in BF16 compiled or in FP32 eager inference\color{blue}{\textbf{Graphene moves the accuracy–cost Pareto frontier forward, whether all methods run in BF16 compiled or in FP32 eager inference}}: it is the most accurate model, and every model that is faster or uses less memory is at least 0.45 dB lower on Urban100 ×4.

Revised manuscript. The main paper and supplementary are now one PDF. Figure 1 and Tables 3, 4, 13 and 14 now measure every method in BF16 compiled inference, and Appendix E.3 repeats every measurement in FP32 eager inference (Table 15, Figure 8, Table 16). Appendix E.1–E.2 add the crop-size and training-data results. Changed figures, tables and new sections are in blue.\color{blue}{\textbf{Changed figures, tables and new sections are in blue.}}


Accuracy–cost frontier.\color{blue}{\textbf{Accuracy–cost frontier.}} Every latency and peak-memory value is measured with one script on one GH200 for every method\color{blue}{\textbf{Every latency and peak-memory value is measured with one script on one GH200 for every method}}, in BF16 compiled inference (BF16 autocast, torch.compile max-autotune, TF32 off; Section 4.1) and in FP32 eager inference (TF32 off; Appendix E.3). Classical ×4 at 64×64 LR input, every method in Table 3 published since 2024:

Method (venue)Params (M)FLOPs (G)BF16: latency (ms) / memory (MiB)FP32: latency (ms) / memory (MiB)Urban100 (dB)
IPG (CVPR 2024)17.0268.511.31 / 394.861.4 / 522.928.13
ATD (CVPR 2024)22.0263.59.00 / 123.064.2 / 222.828.22
MambaIR (ECCV 2024)20.6226.044.76 / 214.143.3 / 249.827.68
MambaIRv2 (CVPR 2025)23.1228.033.88 / 126.784.7 / 224.627.89
PFT (CVPR 2025)19.8188.5114.99 / 533.2181.6 / 621.828.20
IET (ECCV 2026)19.8248.151.68 / 530.467.7 / 705.928.43
Graphene13.0137.08.34 / 192.654.7 / 278.128.67

The frontier holds at every measured input size\color{blue}{\textbf{The frontier holds at every measured input size}}, with all classical methods included:

×4, Urban100Every faster modelEvery model with less memoryGraphene against IET
BF16, 64×64 (Table 3, Figure 1)≥0.70 dB lower≥0.45 dB lower6.2× faster, 2.75× less memory\color{blue}{\textbf{6.2× faster, 2.75× less memory}}
FP32, 64×64 (Table 15, Figure 8)≥0.70 dB lower≥0.45 dB lower1.24× faster, 2.54× less memory\color{blue}{\textbf{1.24× faster, 2.54× less memory}}
BF16, 320×180 / 1280×720 (Table 4)≥0.45 dB lower≥0.45 dB lower6.3× / 7.3× faster, 3.3× / 3.5× less memory\color{blue}{\textbf{6.3× / 7.3× faster, 3.3× / 3.5× less memory}}
FP32, 320×180 / 1280×720 (Table 16)≥0.45 dB lower≥0.45 dB lower3.0× / 2.9× faster, 3.3× / 3.5× less memory\color{blue}{\textbf{3.0× / 2.9× faster, 3.3× / 3.5× less memory}}

In every setting, Graphene is more accurate, faster and leaner than IET, the strongest prior method, so it lies beyond the previous Pareto front, not on it.\color{blue}{\textbf{In every setting, Graphene is more accurate, faster and leaner than IET, the strongest prior method, so it lies beyond the previous Pareto front, not on it.}} Graphene-light does the same against IET-light, and every faster or leaner lightweight model is at least 0.82 dB lower on Urban100 ×4 in every setting (Tables 3, 4, 14–16). Graphene has the highest PSNR in all 15 classical dataset–scale settings (Table 13); Graphene-light leads 12 of 15 lightweight settings and all five at ×4 (Table 14). Baselines are changed only where their official code cannot run the setting (CAT-A, IPG, PFT, IET; † in Figure 1). Where several options exist, we report the fastest, the one most favorable to the baseline. Section 4.1 (Efficiency) states each change and every option measured, and Appendix E.3 the options used at larger LR inputs. Graphene leads in FP32 eager inference as well:\color{blue}{\textbf{Graphene leads in FP32 eager inference as well:}} at every input size, Graphene and Graphene-light are more accurate, faster and leaner than IET and IET-light, and every faster or leaner model is at least 0.45 dB lower on Urban100 ×4 (Appendix E.3: Table 15 and Figure 8 at 64×64, Table 16 at 320×180 and 1280×720).


The two modules add 0.60 dB; the quoted 0.03–0.04 dB values compare single design choices. Tables 1–2 train every variant from scratch under one matched protocol (DIV2K, 10,000 iterations, matched shared-parameter initialization, minibatches and optimization; Section 4.2):

  • IRB and HistAttn together add 0.60 dB at the same window, 7.5× the six-seed SD of 0.08 dB\color{blue}{\textbf{IRB and HistAttn together add 0.60 dB at the same window, 7.5× the six-seed SD of 0.08 dB}} (30.33 → 30.93 dB; Table 1(a), Table 8).
  • For the same 20.4G FLOPs, IRB adds 0.55 dB and a larger window 0.19 dB\color{blue}{\textbf{For the same 20.4G FLOPs, IRB adds 0.55 dB and a larger window 0.19 dB}} (Table 1(a)).
  • In a second backbone, after 50,000 training iterations, IRB improves HAT on all five ×2 benchmarks, by 0.16–0.94 dB (Table 9).

The 0.03 and 0.04 dB values quoted from Tables 1(b) and 2 compare single design choices at identical or nearly identical cost, not whole modules.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.