{"id":"8274fa88-7e61-4087-a7f1-8e670a3ba956","arxiv_id":"2501.06838","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"GSASR predicts free-form 2D Gaussians from a low-resolution image and rasterizes them at arbitrary scales, outperforming prior arbitrary-scale super-resolution methods on standard benchmarks.","lead":"This paper introduces GSASR, a super-resolution method that represents image details as many small 2D Gaussian shapes predicted from the low-resolution input and renders them at any target size with custom GPU code. It reports better image quality than prior implicit-network methods and, at high zoom factors, faster inference.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Rasterization ratio is stated as 0.1 in §3.1 but as 1.0 in Suppl. §6.3; Table 2 timings match r=0.1, so the speed claim depends on which value the released code actually uses.","rationale":"Reader's verdict CONDITIONAL is appropriate. Among the concerns raised, the rasterization-ratio contradiction is the most load-bearing because it targets the speed half of the central claim and is internally checkable. The tiling/boundary-consistency assumption flagged by the reader is a legitimate unanalyzed generalization risk, but the paper's strong results on Set5/Set14/Urban100/BSDS100 (many with sizes not multiples of k=12) provide indirect empirical evidence that padding or shifting handles it acceptably; a dedicated test would still be valuable. The table typos (e.g., ×4 Meta-SR PSNR 33.36) are copy-paste errors that do not affect GSASR's relative standing. The r=0.1 vs r=1.0 contradiction, by contrast, if resolved in favor of r=1.0, would invalidate Table 2's timings and the speed claim; if resolved in favor of r=0.1, it is a simple but material typo. Because the paper ships code, the check is decisive. We therefore recommend keeping the CONDITIONAL verdict until the code is inspected and the text corrected. No change to the reader's verdict.","tokens_in":46308,"tokens_out":12800,"duration_ms":122992,"concrete_test":"Download the released code/models from https://github.com/ChrisDud0257/GSASR and inspect the default rasterization ratio (search for 'rasterization_ratio' or 'ratio' in the rasterizer/config). Then rerun the Table 2 timing experiment (100 720×720 GT patches from DIV2K, single A100, EDSR backbone) with the default setting and with r forced to 1.0. If the default r reproduces the reported 543 ms (×4) / 195 ms (×8), the speed claim is confirmed and §6.3 is a typo requiring correction; if the default is r=1.0 or the timings match the r=1.0 column, Table 2 is wrong and the efficiency claim should be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central efficiency claim ('much faster than state-of-the-art INR-based models') rests entirely on the rasterization ratio r. Section 3.1 states: 'we set the ratio as r = 0.1'. Supplementary Section 6.3 states: 'To balance between the efficiency and performance, we set r = 1.0 cautiously.' These are contradictory. Supplementary Table 13 lists ×4 inference times of 543 ms for r=0.1 and 9419 ms for r=1.0, and ×8 times of 195 ms vs 3285 ms. Main Table 2 reports GSASR inference times of 543 ms (×4) and 195 ms (×8), exactly matching the r=0.1 columns of Table 13. Therefore the published speed numbers are consistent only with r=0.1. If the released model actually uses r=1.0, the reported times are wrong by a factor of ~17 and the high-scale speed advantage disappears: at ×12, GSASR's 91 ms would become roughly 1500 ms, slower than CiaoSR's 540 ms. If the released model uses r=0.1, the speed claim stands but Supplementary §6.3 is a material error. Either way, the manuscript must state unambiguously which ratio is used, and the code must match. This is an internal inconsistency at the heart of the paper's headline advantage, not a matter of external consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GSASR, a feed-forward 2D Gaussian splatting model for arbitrary-scale super-resolution. An LR image is encoded by a standard SR backbone, Gaussian embeddings are conditioned on the encoded features through window-based cross-attention and Gaussian interaction blocks, and a custom differentiable CUDA rasterizer renders the SR image by sampling the predicted continuous Gaussians at scale-dependent positions. The model is trained with L1 loss on scales sampled from [1,4] and tested on scales up to ×30. The main claims are that GSASR outperforms INR-based ASR methods in quality and is much faster, especially at high scaling factors.","tokens_in":46681,"tokens_out":7794,"duration_ms":64013,"significance":"If substantiated, the paper's central claim is significant: a Gaussian-based feed-forward representation would outperform INR-based arbitrary-scale SR in both fidelity and speed, while scaling efficiently to large upsampling factors. The manuscript's strengths include the breadth of evaluation (multiple datasets, two encoder backbones, four metrics, scales up to ×30), the ablations reported in the supplementary material, the exploration of a larger HAT-L backbone, and the promise of released code and models. The CUDA rasterization is a useful engineering contribution, and the position-distribution visualizations in Figures 6 and 9 give useful insight into why the representation helps on textured regions.","major_comments":[{"comment":"The sentence 'GSASR outperforms existing methods in terms of both fidelity metrics (PSNR and SSIM) and perceptual quality metrics (LPIPS and DISTS) under scaling factors from ×2 to ×30' is not supported by the LSDIR columns of Table 1. At ×30, CiaoSR has higher PSNR (19.87 vs 19.82) and lower LPIPS (0.7284 vs 0.7368); at ×24 and ×18, CiaoSR also has higher PSNR (20.40 vs 20.38 and 21.18 vs 21.14). Please either restrict the claim to the datasets and metrics where it holds, or provide a statistical or qualitative justification for the small differences at the largest scales.","section":"§4.2, Table 1 (LSDIR columns)"},{"comment":"The rasterization ratio is stated as r = 0.1 in §3.1 ('we set the ratio as r = 0.1'), but Supplementary §6.3 concludes 'we set r = 1.0 cautiously.' Supplementary Table 13 reports r = 1.0 inference times of 9419 ms (×4) and 3285 ms (×8), while main-text Table 2 reports 543 ms and 195 ms for GSASR, exactly matching the r = 0.1 columns. This is an internal contradiction on the load-bearing efficiency claim. Please state unambiguously which ratio the released code uses and make the text, tables, and code consistent; if r = 1.0 is the deployed setting, the reported speed advantage over CiaoSR no longer holds.","section":"§3.1 vs Supplementary §6.3"},{"comment":"The introduction's claim that 'GSASR is much faster than state-of-the-art INR-based models' is contradicted by Table 2 at low scales: at ×2 GSASR takes 1573 ms versus 86 ms for LINF, 107 ms for SRNO, 126 ms for LTE, and 454 ms for LIIF; at ×3 GSASR takes 806 ms versus 89 ms for LINF and 114 ms for SRNO. The same pattern appears in Supplementary Table 11 with the RDN backbone (×2: 1679 ms vs 178 ms for Meta-SR and 202 ms for LINF). The speed advantage is real only against some methods and mostly at larger scales (e.g., versus CiaoSR at ×3/×4 and beyond). Please rephrase the speed claims to state this precisely, or report timings that do not depend on the rasterization ratio.","section":"Table 2 and Supplementary Table 11"},{"comment":"The arbitrary-image-size claim relies on duplicating a single learned Ebase across all windows and letting reference positions plus window cross-attention specialize the embeddings. The paper does not analyze whether this tiling produces consistent Gaussians across window boundaries, and training is done on fixed 48×48 patches while testing uses larger images. Please add a quantitative check (e.g., PSNR computed on boundary strips versus image interiors, or a visualization of seams) or at least a discussion of when this tiling assumption could break.","section":"§3.2 Gaussian Embedding"}],"minor_comments":[{"comment":"The conclusion contains typos: 'brandly-new' should be 'brand-new' and 'all thoes ablation studies' should be 'all those ablation studies'.","section":"§5"},{"comment":"The final sentence 'we set r = 1.0 cautiously' directly contradicts the main text and the table in the same section; please reword so the chosen default is clear.","section":"Supplementary §6.3"},{"comment":"The caption lists 'scaling factors × 2, × 3, × 4, × 8, × 8'; the second ×8 appears to be a duplicate, probably intended as ×6 or another scale, and the table rows should be checked against the caption.","section":"Supplementary Table 8 caption"},{"comment":"In the ×4 row for RDN, the LPIPS value '23.88' appears to be a typo for '0.2388' (or a similarly formatted value).","section":"Supplementary Table 11"},{"comment":"The caption says 'differential rasterization' where 'differentiable rasterization' is meant.","section":"Figure 2 caption"},{"comment":"The reference title contains 'manga eetrieval' and should read 'manga retrieval'.","section":"Reference [37]"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is generally well executed and the quality results are extensive, but the internal inconsistency about the rasterization ratio and the overbroad speed and quality claims need to be resolved before publication. I do not doubt the authors' intent to release code; however, the published tables and the stated default must match the released implementation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Fairly solid ASR paper that deserves referee time, but the headline speed claim is shakier than the authors let on. The quality story is credible: GSASR beats the INR-based crowd (CiaoSR, LINF, etc.) and GaussianSR on DIV2K and several other benchmarks, consistently across PSNR/SSIM/LPIPS/DISTS, with a unified Y-channel protocol and the same EDSR/RDN backbones. The gains are modest in absolute terms—about 0.2 dB over CiaoSR at ×4—but they do not flip sign across scales or datasets. Learnable free Gaussian positions, the window-based condition injection, and the custom CUDA scale-aware rasterizer are real, if incremental, contributions over GaussianSR, which the paper cites. Ablations cover the key hyperparameters, and the code and models are promised.\n\nThe efficiency claim needs surgery. The paper's own Table 2 at ×2 and ×3 shows GSASR at 1573 ms and 806 ms, while LINF runs at 86 ms and 89 ms. So \"much faster than state-of-the-art INR-based models\" is only true at ×8 and above, and the intro's phrasing overstates it. More seriously, the rasterization ratio r is given as 0.1 in Section 3.1 but as 1.0 in Supplementary 6.3, which says \"we set r = 1.0 cautiously.\" Table 13 lists inference times for both values; the main-text Table 2 times match the r=0.1 columns exactly. So the published speed numbers are consistent only with r=0.1. If the released model actually uses r=1.0, the times are off by a factor of roughly 17 and the high-scale speed advantage disappears: at ×12, 91 ms becomes about 1500 ms, slower than CiaoSR's 540 ms. If it uses r=0.1, then Supplementary 6.3 is wrong. Either way the manuscript must state unambiguously which ratio is used, and the released code must match. This is not an external-consensus issue; it is an internal contradiction at the core of the main claimed advantage.\n\nA softer concern: arbitrary-size generalization rests on duplicating a shared base Gaussian embedding across windows, and the paper gives no analysis of boundary consistency or scale invariance of the predicted Gaussians. It may work fine empirically, but a referee should ask for at least a discussion or a simple experiment on window-boundary artifacts. There are also a few table typos (duplicated ×8 in Table 8's header, an LPIPS value of 23.88 in Table 4 of the supplement, and copy-pasted numbers in Table 17), which suggest a rushed supplementary pass but do not affect the main results.\n\nWho should read this: anyone working on arbitrary-scale SR or on Gaussian-based image representation. The paper is worth engaging with seriously, but it should not be accepted until the rasterization ratio is clarified and the speed claims are re-scoped to the scales where they actually hold. I would send it to peer review, expecting heavy revision on the efficiency section.","headline":"Solid quality gains and a real high-scale efficiency story, but the paper's speed claim is undercut by a self-contradictory rasterization-ratio setting that must be resolved.","tokens_in":47186,"tokens_out":3296,"would_cite":false,"duration_ms":29697,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A feed-forward network that predicts image-conditioned 2D Gaussians and renders them with a scale-aware CUDA rasterizer can outperform implicit neural representations for arbitrary-scale super-resolution in both quality and speed.","keywords":["2D Gaussian splatting","arbitrary-scale super-resolution","implicit neural representation","scale-aware rasterization","continuous image representation","feed-forward decoding","window attention","differentiable rendering"],"falsifier":"Render the same natural image after partitioning it into windows of size 12 and as a whole, then compare outputs along the seams: if the tiling assumption fails, visible discontinuities or a measurable PSNR gap should appear. A second check is to run GSASR on LR images whose width and height are not multiples of 12 and compare PSNR against divisible sizes.","tokens_in":46129,"feed_emoji":"🖼️","tokens_out":6346,"duration_ms":56642,"temperature":0.7,"pith_summary":"The paper tries to establish that 2D Gaussian splatting can be turned into a generalizable, feed-forward model for arbitrary-scale image super-resolution, and that such a model beats implicit neural representation (INR) based methods on both quality and speed. It proposes predicting image-conditioned 2D Gaussians, each with opacity, position, shape, orientation, and color, from the low-resolution image in one forward pass, then rendering the high-resolution output by sampling these continuous Gaussians with a scale-aware GPU rasterizer. If true, this replaces per-pixel MLP queries with region-level primitives that concentrate on textured areas, giving crisper edges and much faster inference, such as 91 milliseconds versus about 540 milliseconds at scaling factor ×12 for a 720×720 output.","feed_headline":"2D Gaussians outclass INR models for arbitrary-scale super-resolution","feed_subtitle":"Feed-forward rendering of continuous Gaussians beats per-pixel MLP queries on quality and speed, ×2 to ×30.","key_machinery":"The central object is the 2D Gaussian as an image primitive: each Gaussian contributes α·c·f(x,y) where f is a bivariate normal density with mean μ, standard deviations σ, and correlation ρ, and the image is the sum of all Gaussians. Because this sum is a continuous field, the scaling factor only changes the sampling interval, so arbitrary magnifications are rendered by the same parameter-free rasterizer. The second load-bearing piece is the differentiable CUDA rasterizer that processes Gaussians in parallel and touches only pixels within a local range, controlled by a rasterization ratio r = 0.1, reducing complexity to O(r²s²HWN) while keeping gradients flowing to the decoder.","core_discovery":"The paper claims that representing an LR image as a sum of N continuous 2D Gaussians, rather than as a field of individually queried pixels, gives a more powerful and cheaper representation for arbitrary-scale super-resolution. Each Gaussian carries opacity, center position, standard deviation, correlation coefficient, and peak RGB color, so it can locally adapt to the shape and orientation of textures. The decoder conditions learnable Gaussian embeddings on LR image features through windowed cross-attention and shifted-window self-attention, then five heads predict the Gaussian properties. A differentiable, GPU-parallel, scale-aware rasterizer samples the continuous field at a density set by the scaling factor, with a local-range truncation that keeps complexity low. On standard benchmarks with EDSR and RDN backbones, the paper reports that GSASR outperforms listed INR methods and the concurrent GaussianSR in PSNR, SSIM, LPIPS, and DISTS across scaling factors from ×2 to ×30, and that its inference is substantially faster at high scales.","pith_inferences":["If the window-tiling generalization is shown to be safe, the same Gaussian decoder could be applied to other continuous-coordinate tasks, such as image compression or video frame interpolation, without changing the rendering core.","The speed advantage over INR models at high scales suggests that the bottleneck of arbitrary-scale super-resolution is per-pixel MLP sampling rather than the number of Gaussians, so models with fewer, adaptively placed Gaussians may reach similar quality at even larger scales.","A testable extension is to replace the fixed window cross-attention with overlap-aware or global attention for a subset of embeddings, preserving locality where it helps while checking whether boundary consistency improves."],"forward_implications":["A single GSASR model, trained with scaling factors sampled from 1.0 to 4.0, reportedly keeps improving over INR baselines on unseen scales up to ×30, so the Gaussian representation itself carries scale generalization.","Because the rasterizer is differentiable, the whole pipeline of encoder, Gaussian decoder, and rendering is trained end-to-end with L1 loss, so the Gaussians are optimized directly for the final super-resolved image.","At ×12 for a 720×720 output, GSASR reports 91 ms versus about 540 ms for CiaoSR, indicating that windowed attention plus the GPU rasterizer avoids the per-pixel MLP cost that dominates INR models.","The learnable center offsets let Gaussians cluster on textured regions while flat areas stay covered by evenly spaced Gaussians, which is the mechanism behind the reported sharpness on edges and windows."],"supporting_citations":[{"why":"Provides the INR-based local implicit image function baseline for arbitrary-scale SR that GSASR claims to beat.","marker":"[11]"},{"why":"Supplies the Gaussian splatting representation of opacity, mean, covariance, and color that the 2D version here adapts.","marker":"[23]"},{"why":"The state-of-the-art INR model with scale-aware attention that GSASR compares against on both quality and speed.","marker":"[2]"},{"why":"The concurrent 2D Gaussian-based ASR method whose fixed Gaussian positions and PyTorch rasterization GSASR contrasts with.","marker":"[20]"},{"why":"The EDSR backbone used to extract features from the LR input in the main experiments.","marker":"[33]"},{"why":"Provides the windowed and shifted-window self-attention mechanisms used for embedding interaction and condition injection.","marker":"[35]"}],"fun_headline_variants":["GSASR: 2D Gaussian splatting outperforms INR super-resolution","Arbitrary-scale super-resolution with efficient 2D Gaussians","Gaussian splatting speeds up and sharpens any-scale SR","2D Gaussian splatting for super-resolution: any scale, faster"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that arbitrary image sizes work rests on the untested assumption that duplicating one shared base Gaussian embedding across all windows produces Gaussians that stay consistent across window boundaries.","fun_headline_variants_meta":{"raw":{"variants":["GSASR: 2D Gaussian splatting outperforms INR super-resolution","Arbitrary-scale super-resolution with efficient 2D Gaussians","Gaussian splatting speeds up and sharpens any-scale SR","2D Gaussian splatting for super-resolution: any scale, faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000261,"raw_usage":{"total_tokens":1640,"prompt_tokens":1040,"completion_tokens":600,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":656,"completion_tokens_details":{"reasoning_tokens":521}},"tokens_in":656,"tokens_out":600,"duration_ms":5464,"temperature":1.0,"reasoning_tokens":521,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:49:17.088273+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render the same natural image after partitioning it into windows of size 12 and as a whole, then compare outputs along the seams: if the tiling assumption fails, visible discontinuities or a measurable PSNR gap should appear. A second check is to run GSASR on LR images whose width and height are not multiples of 12 and compare PSNR against divisible sizes.","supporting_citations":[{"cited_title":"Meta-sr: A magnification-arbitrary network for super-resolution","cited_arxiv_id":null,"evidence_quote":"Provides the INR-based local implicit image function baseline for arbitrary-scale SR that GSASR claims to beat."},{"cited_title":"Super-resolution neural operator","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian splatting representation of opacity, mean, covariance, and color that the 2D version here adapts."},{"cited_title":"Ciaosr: Continuous implicit attention-in-attention network for arbitrary-scale image super-resolution","cited_arxiv_id":null,"evidence_quote":"The state-of-the-art INR model with scale-aware attention that GSASR compares against on both quality and speed."},{"cited_title":"Sketch-based manga eetrieval using manga109 dataset","cited_arxiv_id":null,"evidence_quote":"The concurrent 2D Gaussian-based ASR method whose fixed Gaussian positions and PyTorch rasterization GSASR contrasts with."}],"review_version":1}