acceptodds
Under review as a conference paper at ICLR 2027

RefPLC: Latent Recovery and State-Conditioned Pixel Modulation for Reference-Based Super-Resolution

Abstract

Severe degradation removes the very cues needed to identify missing image detail. A sharp reference can supply related evidence, but differences in pose, viewpoint, and appearance make direct transfer unreliable. Reference-guided generation must therefore address two linked questions: how to turn an unaligned observation into a target-compatible condition, and how to determine its additional contribution as reconstruction evolves. We propose RefPLC, which treats reference guidance as bounded restoration increments at the latent-condition and RGB-velocity interfaces of pixel diffusion. Restoration-Oriented Latent Transport (ROLT) aligns detail-sensitive VAE differences and interprets them relative to the degraded target, preserving the LQ latent as the condition's residual base. State-Conditioned Velocity Modulation (SCVM) compares shared-decoder responses with and without reference evidence at the same sampling state, then spatially modulates the resulting correction. The two stages connect reference preparation to its use within one restoration trajectory. On Single717, RefPLC achieves 25.81 dB PSNR and 67.89 MUSIQ, improving over the equally adapted PiD baseline by 1.89 dB and 4.02, respectively. External evaluation yields 45.88 FID on WR-SR and 28.17 dB PSNR on Face. In a paired 90-target study, Top references improve PSNR by 0.761 dB over the same trained model without a reference. These results support separating target-conditioned evidence recovery from state-dependent reference application.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
reference-based super-resolution, image restoration, pixel diffusion, latent conditioning, reference-guided generation

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_rJ4SOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes RefPLC for reference-based super-resolution. It combines latent reference recovery with state-conditioned pixel modulation and reports improvements over the adapted PiD baseline on several benchmarks.

Critical strengths

1.The method addresses both reference alignment and state-dependent reference usage.

2.The paper includes external benchmarks, module ablations, and a reference-condition study.

Critical weaknesses
  1. The methodological novelty is limited. ROLT and SCVM mainly combine existing correspondence, residual conditioning, and spatial modulation techniques.

  2. The ablation does not demonstrate clear synergy. SCVM alone achieves better SSIM and LPIPS than the complete model.

  3. The qualitative evidence is weak. In Figure J.9, outputs under different references and Null appear nearly identical.

  4. The computational overhead is unclear. Runtime, memory usage, and FLOPs are not reported despite repeated decoder evaluations.

  5. The internal training set and Single717 benchmark lack sufficient details about data sources, reference selection, and public availability.

Questions for authors
  1. Why does SCVM alone achieve better SSIM and LPIPS than the complete model?

  2. How different are the outputs under Top, Medium, Distant, and Null references?

  3. What are the inference latency, memory usage, and FLOPs of RefPLC?

  4. How are the Top, Medium, and Distant references selected?

  5. What are the data sources of the internal training set and Single717 benchmark?

2 comments

Dear Reviewer rJ4S,

Thank you for your questions. We address each point using the revised manuscript.

[W1] Methodological contribution

RefPLC connects two restoration residuals (Sec. 4; Table I.1). Before PiD's condition projection, ROLT transports sharp-minus-blurred VAE detail, applies a target-conditioned shared-map difference, and writes a gated, RMS-limited latent increment. At the RGB-velocity output, SCVM differences shared-decoder responses to reference and zero evidence at the same state, then applies a learned spatial gain. ROLT's recovered condition determines the base prediction seen by SCVM, connecting evidence preparation to its state-dependent use. Figs. I.1–I.4 trace component effects on signals and GT error; Appendix K measures off/on corrections and learned/unit gains at fixed states.

[Q1] SCVM-only versus Full

Table 4 uses the same training protocol. ROLT changes zLz_L to z~L\widetilde z_L within the same VAE space; SCVM's base prediction and correction consequently depend on this condition. Differences in PSNR (dB)/SSIM/LPIPS-VGG are:

ComparisonΔPSNRΔSSIMΔLPIPS
PiD → ROLT+0.81+0.0176−0.0144
PiD → SCVM+1.76+0.0355−0.0349
ROLT → Full+1.08+0.0107−0.0124

SCVM improves all three metrics under both conditions. SCVM-only has higher SSIM (0.7151 vs. 0.7079) and lower LPIPS (0.4382 vs. 0.4463), while Full gains 0.13 dB PSNR and 2.49 MUSIQ over SCVM-only and leads all eight no-reference metrics among the four configurations. Sec. 5.3 distinguishes these final rankings from the gain of adding SCVM under each condition.

Fig. I.4 fixes weights, target, reference, and noise. On the same 64×6464\times64 arm–grille crop, Full obtains 22.086 dB, exceeding HH bypass, caption scale 1, global-only, and local-only controls by 0.093/0.098/0.080/0.126 dB. Figs. I.1–I.3 trace the individual operations.

[Q2] Reference-dependent output differences

Across 90 targets, Top/Medium/Distant improve mean PSNR over the same model's Null output by 0.7614/0.6380/0.0758 dB; all reduce LPIPS (Tables 3, G.7). Fig. J.9 now shows common enlarged crops, crop PSNR, and GT-error maps.

In J.10, Top improves map-crop PSNR over Null by 5.731 dB, lowers MSE from 119.76 to 32.00 (73.28%), and changes the output by a mean absolute 6.22 8-bit RGB levels. In J.11, Top improves gong crop/full-image PSNR by 2.302/1.042 dB and lowers crop MSE from 134.98 to 79.45 (41.14%). MSE-to-GT measures accuracy; MAE-to-Null locates output changes. Each heatmap row shares a scale across conditions, revealing measurable local changes within similar overall compositions.

[Q3] Computational cost

Table F.2 uses one H20, inference batch 1, 25 steps, guidance 5, BF16 mixed precision, and an FP32 sampler, covering multiple aspect ratios.

Model/scaleMean s/imagePeak GiBCounted TFLOPs/image
PiD/5122.45511.90130.65
Full/5123.13812.42137.64
PiD/2K42.14517.362611.56
Full/2K45.89218.522649.33

At 2K, added latency is 3.747 s (8.89%) and peak memory increases by 1.16 GiB. Pair-level evidence is cached; both SCVM decoder passes share one base prediction at each velocity evaluation. CUDA-synchronized timing includes transfer, preprocessing, text embedding, and restoration; it excludes model loading, offline caption generation, and image I/O. Peak allocated memory includes weights. Counted FLOPs cover convolutions, matrix multiplication, and attention (two operations per multiply–add), excluding normalization and most elementwise operations.

[Q4] Reference tiers

Video candidates are quality-screened frames sampled at 1 fps from the GT's 4K source video; non-video references use action/viewpoint/lighting edits with subject/text checks. Overall-similarity and local-correspondence scores fuse DINOv2/VGG16 with weights 0.7/0.3. Top prioritizes overall similarity, Medium local correspondence, and Distant is the least globally similar member of the selected triplet. GT, LQ, LQ caption, and seed are fixed. Each run uses one reference with its caption; Null omits it. Single717 video targets use Top only (Sec. 3; G.1, G.6).

[Q5] Sources and availability

The datasets derive from purchased image/video resources authorized for AI training, covering people, products, scene text, and video frames. Training contains 97,589 GT–Ref pairs from 64,427 targets; Single717 has 717 targets with one reference each. Neither is currently public. Upon acceptance, we will release the training/test data permitted for redistribution, with pairings, tier annotations, and construction scripts.

Area_Chair_TrTGOpenReview ↗

Reminder for Rebuttal

Dear Reviewer,

The authors have provided their rebuttal. Please leave your comments asap.

Your AC

Reviewer_S3NGOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents RefPLC, a reference-based image super-resolution framework built on a pretrained pixel diffusion model, PiD. It introduces two complementary reference-guidance modules. Restoration-Oriented Latent Transport (ROLT) aligns sharp-minus-blurred VAE features from an unaligned reference image and converts them into a gated, amplitude-bounded residual for the LQ latent condition. State-Conditioned Velocity Modulation (SCVM) uses the difference between a shared decoder’s responses with and without reference evidence to produce a spatially modulated correction to the RGB velocity at each sampling state. Experiments on benchmarks demonstrate competitive restoration performance, while a paired reference-condition study examines the influence of reference similarity.

Critical strengths
  1. Well-motivated formulation. The paper distinguishes between identifying relevant reference content and determining its additional contribution to the current restoration state. This provides a clear rationale for separating latent-condition recovery from state-dependent pixel-space correction.
  2. Explicit control of reference influence. Both modules inject reference information through residual updates with spatial gating and amplitude constraints. Their zero-evidence properties ensure that the corresponding corrections vanish when no evidence is supplied, providing a useful structural safeguard against reference-independent changes in these paths.
  3. Competitive reconstruction and perceptual results. On Single717, RefPLC improves PSNR by 1.89 dB and MUSIQ by 4.02 over the adapted PiD baseline. The external benchmark evaluations provide additional evidence of its applicability across different image domains.
Critical weaknesses
  1. The reference-induced response does not necessarily represent a beneficial correction. SCVM computes the difference between a shared decoder’s responses with and without reference evidence. This difference captures the decoder’s sensitivity to the evidence, but does not inherently indicate an improvement in restoration accuracy. Although the training objective encourages useful corrections, the current experiments provide limited direct evidence that the resulting proposal and spatial gain reliably retain beneficial changes and suppress harmful ones.

  2. The necessity of the multiple design components remains insufficiently established. The framework combines target-conditioned latent mapping, spatial gating, global–local evidence fusion, and caption-based modulation, while the ablation primarily evaluates ROLT and SCVM as complete modules. Moreover, adding ROLT to SCVM increases PSNR only from 25.68 to 25.81 dB and improves MUSIQ, while reducing SSIM and worsening LPIPS. These results warrant a clearer analysis of which components drive the improvements and how the two modules complement each other.

  3. Some key implementation details remain underspecified. Although Appendix B.2 specifies the dimensionality of the learned matching descriptors and their combination with frozen C2 features, the architecture used to extract these descriptors is not sufficiently described. Likewise, Appendix C.2 expresses caption compatibility and evidence modulation through the abstract functions h_text and R, without specifying how the captions are represented, how the compatibility signal is computed and bounded, or how it modifies the visual evidence. These omissions limit reproducibility and make it difficult to assess the role of these components.

  4. About the figure. The text in the figures is too small to read comfortably at the paper’s normal viewing scale. Please increase the font size, preferably to around 10–12 pt at the final rendered size, and use a consistent, legible serif font, such as Times New Roman, for a clearer and more professional presentation.

Questions for authors
  1. Could the authors provide direct evidence that SCVM improves the current prediction? For example, at the same sampling state and with the same ROLT-enhanced condition, how does the GT reconstruction error of the one-step clean estimate change when the SCVM correction is enabled versus disabled? Reporting this comparison at representative sampling steps would help clarify when the correction is beneficial.

  2. Could the authors provide targeted ablations of key internal components, such as ROLT’s target-conditioned mapping H, SCVM’s caption-based modulation, and its global–local evidence fusion? In addition, could they explain why combining ROLT with SCVM improves PSNR and MUSIQ but reduces SSIM and worsens LPIPS relative to SCVM alone?

  3. Could the authors provide more detail about the third point mentioned in above weakness.

1 comment

Dear Reviewer S3NG,

Thank you for the suggestions. We added measurements and implementation details.

[W1/Q1] Does SCVM improve the current prediction?

Appendix K fixes (xt,t)(x_t,t), the ROLT condition, base velocity, and proposal ptp_t, then compares g=0g=0 and g=αtg=\alpha_t:

y^t(g)=Q(xt−t[vtbase+0.25g⊙pt]).\hat y_t(g)=Q(x_t-t[v_t^{\mathrm{base}}+0.25g\odot p_t]).

Here Q(z)=(Π(z)+1)/2Q(z)=(\Pi(z)+1)/2: Π\Pi removes padding and maps to the original GT grid, followed by RGB rescaling. Floating-point estimates are scored before clipping or quantization. Zero-based kk indexes velocity evaluations; time grids depend on canvas area, and each paired comparison shares the same xt,tx_t,t.

Tiger and Mural both improve at all nine recorded evaluations per target (k=0,6,12,16,18,20,22,23,24k=0,6,12,16,18,20,22,23,24); Table K.1 shows four, each with 2/2 targets improving. At k=6k=6, full-image MSE decreases by 10.22%/5.08% for Tiger/Mural; at k=24k=24, by 3.09%/3.40%. Fig. K.1 shows Tiger/Damselfly crop PSNR gains of 1.016/0.528 dB and full-image gains of 0.468/0.134 dB.

Fig. K.2 fixes the same proposal at Damselfly k=16k=16 and compares off/unit/learned gain (g=0,1,αtg=0,1,\alpha_t), retaining the factor 0.25. Crop PSNR is 41.522/41.742/41.973 dB. From unrounded measurements, learned gain improves over off/unit by 0.451/0.230 dB on the crop and 0.134/0.031 dB on the full image.

[W2/Q2] Component contributions and metric tradeoffs

Table I.2 and Figs. I.1–I.4 report fixed-weight inference interventions, paired with the same inputs, captions, noise, and sampling settings. One-step controls share (xt,t)(x_t,t) but compute their own condition-dependent predictions; bypassing HH changes the base prediction. Final scores follow each configuration's complete sampling trajectory.

  • Mapping HH (I.1): bypass retains correspondence, gating, and RMS limiting. At the beard–jacket boundary, tensor differences have region-mean channel RMS 0.1865 before gating and 0.0447 for the actual latent increments. Full gains 0.634/0.246 dB in one-step crop PSNR at k=6,12k=6,12 and 2.594/1.383 dB in final crop/full-image PSNR.
  • Caption scaling (I.2): compared with rtext=1r_{\mathrm{text}}=1, normal scaling gains 0.190/0.356 dB on the hair/cruise-window crops and 0.108/0.068 dB on their full images.
  • Fusion (I.3): on the window-blind crop, learned fusion gains 0.088/0.577 dB over global-only/local-only evidence.
  • Common region (I.4): Full reaches 22.086 dB on the same 64×6464\times64 crop. Gains over HH bypass, unit caption scaling, global only, and local only are 0.093/0.098/0.080/0.126 dB; the in-figure table reports local scores.

For the trained module configurations (Section 5.3; Tables 4/G.5), changes in PSNR/SSIM/LPIPS-VGG are:

  • PiD → ROLT: +0.81 dB/+0.0176/−0.0144;
  • PiD → SCVM: +1.76 dB/+0.0355/−0.0349;
  • ROLT → Full: +1.08 dB/+0.0107/−0.0124.

ROLT updates zLz_L to z~L\widetilde z_L within the same VAE space; SCVM corrects the base prediction under that condition. SCVM improves all three metrics under both conditioning designs. SCVM-only retains higher SSIM (0.7151 vs. 0.7079) and lower LPIPS (0.4382 vs. 0.4463). Full gains 0.13 dB PSNR and 2.49 MUSIQ over SCVM-only and achieves the best values for PSNR and all eight no-reference metrics among the four configurations.

[W3/Q3] Exact implementation

Descriptors (B.2): detached 32-channel VAE latents are normalized by per-position channel RMS (floor 10−310^{-3}). A 3×33\times3 convolution maps 32→128 channels, followed by two residual blocks. Each branch repeats 8-group GroupNorm–SiLU–3×33\times3 convolution twice, with residual scale 0.5. Separate bias-free 128→128 query/key projections are L2-normalized. Learned/C² cosine similarities have weights 0.25/0.75 (B.4).

Text and modulation (C.2, Eqs. C.4–C.7): DAPE-RAM/RAM generate LQ/reference captions. Shared frozen Gemma-2-2b-it with PiD's instruction provides the first and last 299 hidden states (300×2304). Mean pooling over all 300 positions and a shared trainable bias-free 2304→64 projection give embeddings with cosine similarity ss. Learned scalars a,ba,b produce rtext=0.5+0.5σ(as+b)∈(0.5,1)r_{\mathrm{text}}=0.5+0.5\sigma(as+b)\in(0.5,1) before masking. This image-pair scalar scales fused RGB and quarter-resolution 64-channel VGG evidence, followed by target-valid masks. Bias-free 1×11\times1 projections form 16-channel full-resolution and 32-channel half-resolution RGB features; the latter uses 2×22\times2 average pooling first. PiD retains the LQ caption. The coefficient is reused along the trajectory; fusion weight ω\omega and state-dependent gain αt\alpha_t have separate roles.

[W4] Figure readability

Figs. 2–3 use full-width vector diagrams. Captions expand r,ΔzR,v0,λr,\Delta z_R,v_0,\lambda and identify the tensor interfaces; Appendices B/C give implementation details.

Reviewer_6FepOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RefPLC for reference-based super-resolution under severe degradation and reference misalignment. The method augments a frozen Pixel Diffusion backbone through two components. Restoration-Oriented Latent Transport aligns reference evidence to the target latent grid and recovers a bounded residual relative to the low-quality latent. State-Conditioned Velocity Modulation then compares the shared decoder responses with and without reference evidence at the same sampling state and applies a spatially modulated correction to the diffusion velocity. The experiments include the internal Single717 benchmark, four external benchmarks, branch-level ablations, and a 90-target study using Top, Medium, Distant, and Null reference conditions. On Single717, RefPLC reports 25.81 dB PSNR and 67.89 MUSIQ, improving over the equally adapted PiD baseline by 1.89 dB and 4.02, respectively. The paired reference study reports a 0.761 dB PSNR gain for Top references over the same trained model without a reference.

Critical strengths
  1. The paper presents a clear two-stage view of reference guidance. ROLT first converts an unaligned reference into target-conditioned latent evidence, while SCVM determines the additional reference contribution at each sampling state. This separation is intuitive and addresses an important difficulty in reference-based restoration.
  2. The empirical results are promising and reasonably broad. RefPLC improves substantially over the adapted PiD baseline on Single717, is evaluated on four external datasets, and includes both module-level ablations and a fixed-model reference study. The appendix also provides useful details about tensor shapes, losses, training, and metric protocols.
Critical weaknesses
  1. The paper does not provide an operational definition of h_text or explain the implementation of the modulation function R. The authors should explicitly clarify these details to improve the reproducibility of the proposed method.

  2. The paper describes the internal datasets only in terms of their size and categories, and characterizes the Top, Medium, and Distant tiers qualitatively. However, it does not report the dataset sources, data-splitting and near-duplicate control procedures. The authors should provide these details to improve the transparency and reproducibility of the dataset construction process.

Questions for authors
  1. Could the authors provide the exact implementation of the caption-reliability pathway, including the inputs and architecture or formula of h_text, the numerical range of B_text, and how R applies this signal to f?
  2. Could the authors report the average inference time and peak GPU memory for RefPLC and PiD?
  3. Could the authors clarify the sources, licensing and release status, and train-test or near-duplicate controls for the internal datasets? Please also specify the candidate pool, similarity model or metric, and selection rule used to construct the Top, Medium, and Distant reference tiers.
2 comments

Dear Reviewer 6Fep,

Thank you for your suggestions on reproducibility, computation, and data construction.

[W1/Q1] Caption compatibility and evidence modulation

Section 4.3 and Appendix C.2, Eqs. (C.4)-(C.7), specify the implementation. DAPE-RAM and RAM generate LQ and reference captions, respectively. Both captions are encoded using PiD's native instruction and the same frozen Gemma-2-2b-it encoder. We retain the first hidden state and the last 299 states (300 x 2304), average all 300 positions without an additional pooling mask, and apply a shared, trainable, bias-free 2304-to-64 projection. For the projected embeddings eL,eRe_L,e_R,

s=\cos(e_L,e_R),\quad \mathcal B_{\mathrm{text}}=(0.5,1).$$ Here $a,b$ are learned scalars. The range is before masking; finite precision can round to an endpoint. One coefficient is computed per image pair and reused along sampling. After global-local fusion, $\mathcal R$ broadcasts this coefficient over channels and positions to scale RGB detail and quarter-resolution VGG detail. VGG detail is projected to 64 channels before alignment; RGB detail has three channels. With target-valid mask $V$ and average pooling $A_k$ of kernel/stride $k$, the masked tensors are $$\widetilde f_{\mathrm{rgb}}=V\odot(r_{\mathrm{text}}f_{\mathrm{rgb}}),\qquad \widetilde f_4=A_4(V)\odot(r_{\mathrm{text}}f_4).$$ The decoder input $F_R=\mathcal R(f;r_{\mathrm{text}})$ contains these two tensors plus $P_1(\widetilde f_{\mathrm{rgb}})$ at full resolution (16 channels) and $P_2(A_2(\widetilde f_{\mathrm{rgb}}))$ at half resolution (32 channels). $P_1,P_2$ are bias-free $1\times1$ convolutions: the half-resolution path pools before projection. PiD retains the LQ caption as its native text condition; the reference caption is used for SCVM compatibility. ### [Q2] Inference latency, memory, and FLOPs Table F.2 measures both methods on one NVIDIA H20: inference batch size 1, 25 steps, guidance scale 5, BF16 mixed precision, and an FP32 sampler. Each output scale covers multiple aspect ratios. | Model | Scale | Mean s/image | Peak GiB | Counted TFLOPs/image | |---|---|---:|---:|---:| | PiD | 512 | 2.455 | 11.90 | 130.65 | | RefPLC | 512 | 3.138 | 12.42 | 137.64 | | PiD | 2K | 42.145 | 17.36 | 2611.56 | | RefPLC | 2K | 45.892 | 18.52 | 2649.33 | The added latency is 0.683 s (27.82%) at 512 and 3.747 s (8.89%) at 2K; peak memory increases by 0.52 and 1.16 GiB. CUDA-synchronized timing includes input transfer, preprocessing, text embedding, and restoration; it excludes model loading, offline caption generation, and image I/O. Memory is peak allocated memory, including resident weights. FLOPs count convolutions, matrix multiplications, and attention, with two operations per multiply-add; normalization and most elementwise operations are omitted. ROLT and static reference preparation are computed once per image pair. At each velocity evaluation, SCVM's two lightweight shared-decoder forwards use the same PiD base prediction, followed by the gain head. Eq. (F.1) separates cached and step-dependent computation. ### [W2/Q3] Data sources, availability, and reference tiers The training set and Single717 were constructed from purchased image/video resources licensed for AI training, covering people, products, scene text, and video frames. Training uses 97,589 GT-reference pairs from 64,427 distinct targets; Single717 has 717 targets with one reference each. A separate study evaluates four reference conditions on 90 targets. These internal datasets are not currently public. Upon acceptance, we will release the portions permitted for redistribution, with pairing metadata, tier annotations, and construction scripts. We extract 4K video frames at 1 fps, screen quality, and select candidates at different times in the GT's video. Non-video references use action, viewpoint, or illumination edits, followed by subject/text consistency checks. Each GT and all its references stay in one split. MobileNetV3-Small features screen highly similar GT frames within videos. Cross-split pixel and feature-neighbor screening is followed by verification; confirmed duplicate/near-duplicate training GTs and all their pairs are removed. Candidate matching combines DINOv2 and VGG16 scores with weights 0.7/0.3. Top prioritizes overall similarity, Medium local correspondence, and Distant is the least globally similar member of the selected related triplet. The target, LQ, LQ caption, and seed stay fixed across conditions; Null omits the reference. Section 3 and Appendix G.1 describe task organization; G.4-G.6 specify metric and paired-input conventions.

Thank you for providing the methodological clarifications, efficiency measurements, and additional information on the data construction process. This addresses my remaining questions about implementation, computational cost, and reference-tier construction. The added evidence strengthens the reproducibility and practical assessment of the method. Good luck!

Comments on the paper

1 comment

Dear Reviewers and Area Chair,

Thank you for your careful reading and constructive feedback. We have uploaded a revised manuscript with the supplementary material included after the references. The main text is within the 10-page rebuttal limit.

The main revisions are:

  1. Implementation details and presentation (Section 4, Appendices B-C, and Figs. 2-3): we specify the descriptor network, caption encoding, text compatibility coefficient, and multi-scale evidence modulation, and improve the method diagrams and captions.

  2. Component analysis (Section 5.3, Table G.5, Table I.2, and Figs. I.1-I.4): we clarify the module-ablation comparisons and metric rankings, and add fixed-weight interventions on the latent mapping, caption scaling, and global-local fusion. Fig. I.4 compares the interventions on a common target crop.

  3. Same-state SCVM measurements (Appendix K, Table K.1, and Figs. K.1-K.2): we report paired GT errors with SCVM off/on at fixed sampling states, and compare off, unit, and learned gains for the same correction proposal.

  4. Reference-condition visualizations (Figs. J.9-J.11): we provide common enlarged crops, local PSNR, GT-error maps, and output-difference maps relative to Null.

  5. Computational cost and evaluation protocols (Table F.2, Section 3, and Appendix G): we report inference latency, peak allocated memory, and counted FLOPs under matched settings, and clarify dataset organization and paired evaluation conventions. Additional details on data construction and availability are provided in our reviewer responses.

All section, table, and figure references in our responses refer to this revised PDF. Our point-by-point replies are posted in the corresponding reviewer threads.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.