{"id":"b92002c6-4bd8-44d6-8fc9-8ed3775708dd","arxiv_id":"2505.02175","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A feed-forward model regressing 2D Gaussian splat parameters from three views reports the best Chamfer distance on DTU sparse reconstruction in its comparison table, and competitive novel view synthesis, at roughly 80 times the speed of implicit generalizable baselines.","lead":"A new method turns just three photos of an object into a 3D model and new views, using flat 'surface elements' instead of the more common 3D blobs. It claims top accuracy on a standard benchmark and runs far faster than earlier generalizable methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA reconstruction claim rests on the authors' own UfoRecon reproduction (1.05), which they admit does not match the original; if the original mean Chamfer is below 1.04, SparSplat is not SOTA.","rationale":"The most load-bearing condition for the paper's headline is that its strongest comparison baseline, UfoRecon, is accurately represented. The paper itself says UfoRecon* in Table 1 is its own reproduction and that it could not reproduce UfoRecon's published Chamfer metrics. Since the claimed SOTA is 1.04 versus 1.05, the entire advantage can disappear if UfoRecon's original result is 1.04 or better. This is not a stylistic concern; it is a falsifiable condition on the central claim. I also note the NVS claim has a similar fragility in Table 2: SSIM drops by 0.025 relative to MVSGaussian while PSNR improves by 0.12, so 'state-of-the-art NVS' relies on a PSNR-only reading. The method itself is plausible and has real positive evidence: it adapts MVSGaussian to 2DGS, adds frozen foundation features with a clear ablation (1.17 to 1.04 mean Chamfer), reports a large speed advantage, and evaluates on three datasets. Those contributions remain valuable even if the SOTA phrasing is softened. The reader's weakest assumption correctly identified the UfoRecon baseline, so I agree with the conditional verdict.","tokens_in":19197,"tokens_out":11523,"duration_ms":140273,"concrete_test":"Obtain UfoRecon's original per-scan Chamfer values (from their paper, repository, or authors) and recompute the mean over the exact 15 DTU scans and two 3-view sets used in Table 1; if the original mean is at or below 1.04, the abstract's reconstruction-SOTA claim is false and must be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that SparSplat attains state-of-the-art DTU sparse reconstruction by Chamfer distance (Table 1, mean 1.04). The closest generalizable baseline, UfoRecon, is listed at 1.05, but the footnote and Section 4.1 state that this value comes from running the pretrained model and that the authors could not reproduce the Chamfer metrics originally reported by UfoRecon. The SOTA margin is therefore not measured against the baseline's own reported performance; if the original UfoRecon number is lower than 1.04, the reconstruction claim fails. The margin is also extremely thin (0.01 on the mean), and while the authors averaged UfoRecon over multiple runs, they report no variance or significance testing for either method, so even the reproduced comparison could be within noise. A second, related weakness is the NVS claim: Table 2 shows SSIM 0.938 versus MVSGaussian 0.963, a drop far larger than the 0.12 dB PSNR gain, so 'state-of-the-art NVS' depends on a PSNR-only reading of the metrics. These concerns do not invalidate the method's engineering value, but they mean the headline SOTA claims are not yet secure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SparSplat, a feed-forward, generalizable 2D Gaussian splatting model for joint sparse-view 3D reconstruction and novel view synthesis. Given three posed images, the model predicts pixel-aligned 2DGS parameters by extending the MVSGaussian architecture with features from DINOv2 and MASt3R, and trains with RGB, depth, depth-distortion, and normal-consistency losses. The authors report the best mean Chamfer distance on the DTU sparse reconstruction benchmark (1.04 vs. 1.05 for their reproduced UfoRecon baseline), improve PSNR and LPIPS over MVSGaussian on DTU NVS, demonstrate generalization to BlendedMVS and Tanks and Temples, and highlight an inference time of about 0.8 s versus tens of seconds for implicit generalizable methods. The core technical contribution is a fast generalizable 2DGS pipeline, and the paper argues that the representation and the additional foundation-model features drive the gains.","tokens_in":19496,"tokens_out":3553,"duration_ms":41288,"significance":"If the reported numbers hold, the paper makes a useful contribution: it is, to my knowledge, the first generalizable feed-forward 2D Gaussian splatting method for joint reconstruction and NVS, and it demonstrates a large practical speed advantage over implicit generalizable reconstruction baselines. The ablation showing that MASt3R features improve reconstruction over FPN or DINOv2 features (Table 3) is informative, and the use of dense depth supervision with 2DGS in a generalizable setting is a sensible extension. The methodological core is plausible and the experimental setup is mostly aligned with prior work. However, the two headline claims of state-of-the-art reconstruction and state-of-the-art NVS are not yet secure because of the baseline-reproduction issue and the selective use of metrics, as detailed below.","major_comments":[{"comment":"The headline SOTA reconstruction claim rests on the authors' reproduced UfoRecon mean Chamfer of 1.05, yet the text explicitly states that the authors could not reproduce the Chamfer metrics originally reported in the UfoRecon paper. If the original UfoRecon value is below 1.04, SparSplat is not SOTA on DTU reconstruction. The margin is 0.01 on a mean over 15 scans, and no variance or significance testing is reported for either method, so even the comparison against the reproduced number may be within noise. Please report the originally published UfoRecon values alongside the repro-duced ones, provide per-scan numbers and standard errors for both methods, and either support or temper the abstract's SOTA claim accordingly.","section":"Section 4.1 and Table 1"},{"comment":"The claim of state-of-the-art NVS is metric-selective. Compared with MVSGaussian, SparSplat improves PSNR by 0.12 dB and LPIPS by 0.003, but SSIM drops from 0.963 to 0.938, a relative decrease of about 2.6%. The abstract and Section 4.3 assert SOTA NVS based on PSNR/LPIPS. Please report per-scene metrics and discuss the SSIM drop, or revise the claim to state explicitly on which metrics the method is SOTA and acknowledge the SSIM regression.","section":"Table 2 and Section 4.3"}],"minor_comments":[{"comment":"The caption cites UfoRecon as reference [69] and writes 'UfoRecon*1 [69]', but the bibliography lists UfoRecon as [54]; the footnote marker '1' also appears to dangle. Please correct the citation and place the footnote marker consistently.","section":"Table 1 caption"},{"comment":"The test scan IDs for DTU surface reconstruction are listed as '24, 37, 40, 55, 63, 192, 65, 69, 83, 97, 105, 106, 110, 114, 118 and 122', which contains 16 entries and includes scan 192, while Table 1 and the main text use the standard 15 scans without 192. Please correct this inconsistency.","section":"Supplementary Section C"},{"comment":"The captions contain the typo 'datatset' instead of 'dataset' (repeated in Figures 2, 3, and 4).","section":"Figures 2-4 captions"},{"comment":"The loss weights (lambda_s, lambda_p, lambda_alpha, lambda_beta, lambda_gamma, lambda_1, lambda_2) are only given in the supplementary material; consider moving them to the main text so the objective is self-contained.","section":"Equations (10)-(12)"},{"comment":"There are typos such as 'as-well' in the abstract and 'post-hock' in the introduction; also, in Section 3.3 the text says 'We run master on all pairs' where 'MASt3R' is intended.","section":"Abstract and Section 2"},{"comment":"In the sentence 'Without 2DGS ... the method fails to obtain conherent surfaces', 'conherent' should be 'coherent'.","section":"Section 4.4"}],"recommendation":"major_revision","confidential_remarks":"The UfoRecon baseline discrepancy is the key risk: the authors should contact the UfoRecon authors or otherwise resolve whether the originally reported Chamfer value is below 1.04 before the SOTA reconstruction claim can stand. The self-citation pattern (GeoTransfer [32], Sparfels [33]) is noticeable but does not affect the technical content. The paper is within the scope of the journal and the core method is promising; the required changes are additional experimental reporting and a more careful framing of the SOTA claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: SparSplat is a sensible and reasonably honest engineering contribution — the first feed-forward 2DGS regression I'm aware of — and the speed advantage over implicit generalizable methods is real and useful. The SOTA framing is the weak point: the reconstruction lead over UfoRecon is 0.01 Chamfer on a baseline value the authors had to reproduce themselves, and the NVS claim conveniently ignores an SSIM drop from 0.963 to 0.938.\n\nWhat's new: the paper takes MVSGaussian's architecture, swaps 3DGS primitives for 2DGS, and injects MASt3R dense-pairwise features into the FPN and cost volume. The combination looks new. The ablations earn credit: MASt3R features improve mean Chamfer from 1.17 to 1.04, and adding the depth loss improves it from 4.37 to 1.04 — a 71.8% jump, which suggests the loss is doing the heavy lifting. They also report a genuine 0.8s inference time versus 30-66s for implicit baselines. Those are real contributions.\n\nThe soft spots are proportionate. First, the reconstruction SOTA claim depends on the UfoRecon baseline being 1.05, which is the authors' own reproduction using the pretrained model; they state they could not reproduce the metrics in the original UfoRecon paper. If the original number is below 1.04, the claim fails. There are also no error bars or significance tests on a 0.01 margin, which is likely within run-to-run noise. Second, Table 2 shows PSNR improving by 0.12 dB over MVSGaussian while SSIM drops by 0.025 — a much larger relative change. 'SOTA NVS' is a PSNR-selective reading. Third, generalization results on BlendedMVS and Tanks and Temples are qualitative only, and no code or models are released, which makes verification harder.\n\nTo be clear, these issues don't invalidate the method or the direction. The paper is open about the UfoRecon reproduction problem, the ablations are informative, and the speed result is solid. Who is this for? Anyone working on generalizable Gaussian splatting or fast sparse-view reconstruction; it's a good baseline to know and compare against.\n\nI'd accept it for peer review. It deserves referee time, and the right revision can fix the claims by adding variance, reporting the original UfoRecon value for context, and presenting the NVS comparison without cherry-picking.","headline":"Useful engineering contribution with a real speed win, but the SOTA claims rest on a thin margin and a non-reproduced baseline value.","tokens_in":20076,"tokens_out":2593,"would_cite":true,"duration_ms":28512,"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":"SparSplat claims state-of-the-art sparse-view 3D reconstruction and novel view synthesis by regressing 2D Gaussian surface elements from three input images in one feed-forward pass, at roughly 0.8 seconds per scene.","keywords":["2D Gaussian Splatting","sparse-view 3D reconstruction","novel view synthesis","multi-view stereo","feed-forward prediction","foundation model features","surface reconstruction","DTU benchmark"],"falsifier":"Run the released pretrained model of the previous best implicit reconstruction method on the same 15 DTU test scenes with the same two three-view sets, the same masks, and the same voxel size, and compute the mean Chamfer distance; if it comes out below 1.04, the paper's main accuracy claim is overturned.","tokens_in":18989,"feed_emoji":"🧊","tokens_out":7892,"duration_ms":87938,"temperature":0.7,"pith_summary":"The paper proposes SparSplat, a feed-forward network that takes three posed images and predicts a set of flat 2D Gaussian surface elements for a target view. These elements are rendered into color and depth via Gaussian splatting and fused with TSDF into a 3D mesh, so the same prediction serves both novel view synthesis and surface reconstruction. The authors claim this is the first generalizable feed-forward approach built on 2D Gaussian splatting, and that it reaches a mean Chamfer distance of 1.04 on the DTU sparse-view benchmark, narrowly ahead of the 1.05 reproduced for the previous leading implicit reconstruction method. They also report improved novel view synthesis over the Gaussian-splatting backbone and inference time of about 0.8 seconds, roughly two orders of magnitude faster than implicit baselines. If correct, this makes sparse-view reconstruction practical for interactive applications such as robotics, augmented reality, and quick scanning.","feed_headline":"SparSplat rebuilds 3D scenes from three views in under a second","feed_subtitle":"Sparse-view 3D reconstruction and novel view synthesis in one forward pass, beating slower implicit methods on accuracy.","key_machinery":"The load-bearing object is the 2D Gaussian surface element: a flat, elliptical primitive defined by a center, two tangent scaling factors, a rotation, an opacity, and a color, rasterized by intersecting each pixel ray with the splat's plane. Its role is to make predicted depth consistent across views, so that fusing several rendered depth maps with TSDF yields a coherent surface instead of the warped or fragmented result that 3D Gaussian primitives produce. The network combines a multi-view stereo cost volume that produces depth, a second branch that regresses the remaining attributes, and injected dense pairwise matching features that materially improve the cost volume. Training is end-to-end with color, SSIM, perceptual, depth, depth-distortion, and normal-consistency losses.","core_discovery":"The central claim is that representing predicted scene geometry as flat 2D Gaussian surface elements, rather than volumetric 3D Gaussians, lets a single generalizable feed-forward network solve sparse-view novel view synthesis and 3D mesh reconstruction together. The network predicts per-pixel surface element parameters for a target view from three source views, places each element's center by unprojecting a multi-view stereo depth prediction, and renders with perspective-accurate splatting. On the DTU sparse reconstruction benchmark the authors report a mean Chamfer distance of 1.04, narrowly ahead of the reproduced 1.05 of the previous leading generalizable implicit method, while improving novel view PSNR over the Gaussian-splatting backbone and cutting inference time from tens of seconds to about 0.8 seconds. The paper also shows that enriching the encoder with dense pairwise correspondence features from a pretrained 3D foundation model gives a larger reconstruction boost than monocular semantic features.","pith_inferences":["If the speed and accuracy claims hold beyond the reported benchmarks, sparse-view reconstruction may shift from offline optimization to interactive settings, with pose estimation and feature extraction becoming the main bottlenecks.","A natural testable extension is to replace TSDF fusion with a learned surface extraction directly from the 2D Gaussian primitives, which could preserve detail lost by voxel fusion.","The relative gain from dense correspondence features over semantic features suggests stereo matching quality, not semantic understanding, is the limiting factor in sparse-view reconstruction; even richer correspondence features could push accuracy further.","Because the model builds on an existing multi-view stereo Gaussian backbone, a fair comparison against a stronger version of that backbone would clarify how much of the gain comes from the 2D primitive representation itself."],"forward_implications":["Three-view reconstruction becomes a single forward pass, making mesh extraction and novel view synthesis available in under a second on a single GPU.","A single model can serve both reconstruction and rendering, eliminating the need for separate pipelines for these two tasks.","Because the model generalizes without fine-tuning to unseen indoor and outdoor scenes, it can be applied directly to new captures.","The finding that pairwise 3D foundation features help more than monocular features gives a transferable recipe for improving other multi-view stereo based Gaussian prediction pipelines."],"supporting_citations":[{"why":"Supplies the multi-view stereo backbone, hybrid rendering, and multi-stage training procedure that the model adapts; also the novel-view-synthesis baseline it must beat.","marker":"[48]"},{"why":"Defines the planar 2D Gaussian primitive and perspective-accurate rasterization that yield view-consistent depth for mesh extraction.","marker":"[28]"},{"why":"Provides the dense pairwise correspondence features whose injection into the encoder produces the reported reconstruction gains.","marker":"[41]"},{"why":"The previous leading generalizable implicit reconstruction baseline whose reproduced mean Chamfer distance is the number the paper must beat.","marker":"[54]"},{"why":"Defines the DTU sparse-view test split, the three-view protocol, and the TSDF fusion evaluation used for the reconstruction comparison.","marker":"[50]"},{"why":"Supplies monocular semantic features used in the ablation showing they give only a small gain, unlike the multi-view features.","marker":"[59]"},{"why":"Motivates the strategy of integrating pretrained foundation-model features into a multi-view stereo pipeline.","marker":"[6]"}],"fun_headline_variants":["SparSplat: sparse-view 3D in ~0.8s with 2D Gaussian splatting","One forward pass yields sparse 3D and novel views via 2D Gaussians","SparSplat beats implicit methods in sparse-view 3D speed and accuracy","SparSplat: 2D Gaussians deliver sparse 3D and novel views in 0.8s"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The state-of-the-art reconstruction claim rests on the comparison value used for the previous best method: the authors could not reproduce the lower error that method originally reported, so if that original number is correct, this method may not actually be the most accurate.","fun_headline_variants_meta":{"raw":{"variants":["SparSplat: sparse-view 3D in ~0.8s with 2D Gaussian splatting","One forward pass yields sparse 3D and novel views via 2D Gaussians","SparSplat beats implicit methods in sparse-view 3D speed and accuracy","SparSplat: 2D Gaussians deliver sparse 3D and novel views in 0.8s"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001153,"raw_usage":{"total_tokens":4833,"prompt_tokens":1056,"completion_tokens":3777,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":672,"completion_tokens_details":{"reasoning_tokens":3682}},"tokens_in":672,"tokens_out":3777,"duration_ms":28606,"temperature":1.0,"reasoning_tokens":3682,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:59:20.187152+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released pretrained model of the previous best implicit reconstruction method on the same 15 DTU test scenes with the same two three-view sets, the same masks, and the same voxel size, and compute the mean Chamfer distance; if it comes out below 1.04, the paper's main accuracy claim is overturned.","supporting_citations":[{"cited_title":"2d gaussian splatting for geometrically ac- curate radiance fields","cited_arxiv_id":null,"evidence_quote":"Defines the planar 2D Gaussian primitive and perspective-accurate rasterization that yield view-consistent depth for mesh extraction."},{"cited_title":"Uforecon: Generalizable sparse-view surface reconstruction from arbitrary and unfavorable sets","cited_arxiv_id":null,"evidence_quote":"The previous leading generalizable implicit reconstruction baseline whose reproduced mean Chamfer distance is the number the paper must beat."},{"cited_title":"Sparseneus: Fast generalizable neural sur- face reconstruction from sparse views","cited_arxiv_id":null,"evidence_quote":"Defines the DTU sparse-view test split, the three-view protocol, and the TSDF fusion evaluation used for the reconstruction comparison."}],"review_version":1}