{"id":"4e6f2192-74da-4ada-ad2a-2ab03d912295","arxiv_id":"2607.20417","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ATSplat reconstructs 3D scenes from multiple photos in a single forward pass with ~5.7× fewer Gaussians than dense feed-forward 3DGS and comparable or better rendering quality.","lead":"ATSplat is a fast method for turning multiple photos into a 3D scene, using far fewer 3D Gaussians than older feed-forward methods while keeping image quality. It matters because real-time 3D reconstruction and rendering could become cheaper for AR/VR, robotics, and video-based content creation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"State-of-the-art claim is contradicted by the paper's own RealEstate10K table (iLRM: 28.65/0.900 vs. Ours: 28.46/0.894) and the high-resolution DepthSplat baseline is run off its training resolution; the central 'SOTA with 5.7x fewer Gaussians' claim needs narrowing.","rationale":"The reader's weakest_assumption focuses on uncertainty-score calibration, which is a plausible concern but not the most load-bearing one for the paper's headline claim. The ablation in Table 5 already shows uncertainty-guided selection beats random and FPS selection, so even imperfect calibration still delivers a measurable gain; perfect calibration is not required for the compactness result. The sharper problem is that the headline claim has two parts—'state-of-the-art rendering quality' and '5.7x fewer Gaussians'—and the first part is contradicted by the paper's own RealEstate10K table. Additionally, the high-resolution comparison that would support SOTA at 512x960 uses DepthSplat at a resolution its public checkpoint was not trained for, making the baseline comparison unfair. This is an internal inconsistency rather than a disagreement with consensus, and it directly affects what the paper claims to have shown. The reader's rationale already notes the RealEstate10K contradiction and the high-resolution baseline mismatch, so the reader and I partially agree, but the reader's formal weakest_assumption is a different, less central issue. I would keep the verdict at CONDITIONAL: the method appears plausible and the efficiency result is meaningful, but the abstract's SOTA claim needs to be narrowed and the high-resolution baseline made fair before acceptance.","tokens_in":18999,"tokens_out":6677,"duration_ms":88421,"concrete_test":"Use the released iLRM (and DepthSplat, if available) checkpoints to reproduce Table 1 and Table 3 under exactly the paper's evaluation splits and resolutions; for the high-resolution table, first fine-tune DepthSplat at 512x960 or use a checkpoint trained at that resolution. Compute paired per-scene PSNR/SSIM/LPIPS. If iLRM remains ahead on RealEstate10K and/or the properly matched DepthSplat closes the Table 3 gap, the abstract's 'state-of-the-art rendering quality' claim should be revised to a narrower claim, e.g., 'state-of-the-art among compact feed-forward methods on DL3DV at 256x448.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central contribution is compactness without quality loss: 'state-of-the-art rendering quality while reducing Gaussians by >5.7x.' The efficiency part is credible: the token/anchor design and ablations support adaptive allocation, and the Gaussian-count reductions are real. The SOTA-quality part, however, is not consistently supported. In Table 1 (RealEstate10K, 2 views), iLRM scores 28.65 PSNR / 0.900 SSIM / 0.110 LPIPS while ATSplat scores 28.46 / 0.894 / 0.118; ATSplat is not SOTA on this dataset, and Sec. 7.2 quietly says 'comparable' rather than SOTA. The abstract and conclusion nevertheless claim SOTA across the two datasets. The only high-resolution SOTA evidence (Table 3) is further weakened by a protocol mismatch: DepthSplat's public checkpoint was trained at 448x768 and is evaluated at 512x960, so its 21.33 PSNR is not a fair feed-forward baseline; iLRM at 24.35 is within 0.5 dB. The 5.7x claim is also computed against the densest baselines (131K/688K), not against iLRM (172K/1474K), where the reduction is about 4.7x at high resolution. If the SOTA claim is dropped or narrowed to DL3DV at 256x448, the central scientific message still holds, but as written the abstract overstates the results.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"ATSplat is a feed-forward 3D Gaussian Splatting method that replaces pixel-aligned, per-pixel Gaussian prediction with a sparse set of 3D anchor tokens derived from coarse patch features and predicted depths. An Adaptive Token Expansion (ATE) module predicts per-token uncertainty scores, supervised by D-SSIM error maps of the model's own intermediate renderings, and selectively expands high-uncertainty tokens into additional Gaussians. The method is evaluated on RealEstate10K and DL3DV under standard protocols, including a high-resolution setting, and the paper claims state-of-the-art rendering quality with more than 5.7× fewer Gaussians than dense feed-forward baselines. The central technical contribution is the sparse-to-adaptive allocation that decouples primitive placement from the input image grid.","tokens_in":19454,"tokens_out":4578,"duration_ms":40885,"significance":"If the claims are substantiated, the work makes a useful contribution to feed-forward 3DGS by showing that adaptive capacity allocation — a key strength of per-scene optimized 3DGS — can be restored in a single forward pass. The proposed anchor-plus-offset design and the ATE module are clearly motivated, and the ablations (Tables 4, 5 and supplementary C) demonstrate that both components contribute to the reported gains. The paper also provides useful additional experiments: dynamic Gaussian budgets via uncertainty thresholds (Table 12), zero-shot transfer (Table 9), extrapolation (Table 10), and a unified model across datasets/view counts (Table 11). These strengthen the empirical case that the representation budget can be made scene-dependent rather than resolution-dependent. However, the headline claim that ATSplat achieves 'state-of-the-art rendering quality' is not consistently supported by the paper's own tables, and the high-resolution baseline comparison includes a protocol mismatch. The core efficiency idea is valuable, but the claims need to be narrowed and qualified.","major_comments":[{"comment":"The abstract and conclusion claim 'state-of-the-art rendering quality' with 5.7× fewer Gaussians, but Table 1 on RealEstate10K shows iLRM achieves higher PSNR (28.65 vs 28.46), higher SSIM (0.900 vs 0.894), and lower LPIPS (0.110 vs 0.118). ATSplat is not state-of-the-art on this dataset. Section 7.2 carefully says 'comparable to state-of-the-art feed-forward methods,' which is accurate, but the abstract and conclusion do not match. The claim should be narrowed to DL3DV or to 'state-of-the-art among compact/sub-40K-Gaussian methods.' This is a load-bearing overstatement of the central result.","section":"Abstract and Sec. 7.2, Table 1"},{"comment":"The high-resolution DL3DV comparison includes DepthSplat evaluated at 512×960 using a checkpoint trained at 448×768. This is an off-protocol evaluation and makes DepthSplat's 21.33 PSNR not a fair feed-forward baseline; the gap to ATSplat (24.85) is thereby inflated. The paper acknowledges this in the caption, but the abstract and conclusion still rely on the resulting 'state-of-the-art' claim. I recommend either evaluating at DepthSplat's native resolution (with appropriate scaling) or clearly labeling the number as an out-of-distribution point and excluding it from SOTA claims. Also, the 5.7× reduction figure is computed against the densest pixel-aligned baselines (131K/688K); against iLRM the reduction is roughly 4.7× at high resolution (1474K/311K) and only 1.4× in the DL3DV 6-view setting (172K/120K). The quantitative claim should be stated relative to a specific baseline class.","section":"Sec. 7.3, Table 3"},{"comment":"The ATE module and the central efficiency claim depend on the learned uncertainty score u_i being a reliable proxy for actual rendering error on unseen scenes. The paper provides only qualitative visualizations (Fig. 4) and an ablation (Table 5) showing uncertainty-guided selection beats random selection. No quantitative calibration is reported, nor any analysis of failure cases or sensitivity to the fixed selection ratios ρ_l. Given the method's entire premise is that capacity is allocated to high-error regions, I ask for a quantitative validation, e.g., rank correlation between predicted uncertainty and D-SSIM, or an AUROC-style measure, and an ablation over selection ratios (the supplementary's threshold-based dynamic budgets partially address the latter, but not the calibration). This would substantiate the robustness of the adaptive mechanism.","section":"Sec. 5, Eq. (8); Sec. 7.4, Fig. 4"},{"comment":"The paper's own supplementary shows ATSplat falls behind pixel-aligned baselines in the large-overlap regime (e.g., PSNR 29.13 vs iLRM's 29.43, DepthSplat's 28.32). This is a legitimate limitation, but it is not mentioned in the main paper's discussion or conclusion. Since the paper claims general superiority of the anchor-offset design, this domain-dependent behavior should be disclosed and discussed in the main text, especially because it affects the interpretation of where the method is applicable.","section":"Supplementary Sec. B.1, Table 7"}],"minor_comments":[{"comment":"The phrase 'comparable to state-of-the-art feed-forward methods' is accurate, but the following sentence says '5.7× reduction compared to dense pixel-aligned formulations' — the comparison class is not stated. Please specify that the reduction is against pixelSplat/MVSplat/DepthSplat, not against iLRM.","section":"Sec. 7.2, RealEstate10K paragraph"},{"comment":"The abstract claims real-time rendering at 1136 FPS (512×960), but no table or experiment in the main text measures rendering FPS. Please provide the measurement protocol (GPU, rasterizer, batch size) or remove the number from the abstract if it is not reproducible from the reported experiments.","section":"Abstract and Fig. 1"},{"comment":"The uncertainty loss uses L1 between predicted and stop-gradiated D-SSIM error maps. The choice of D-SSIM vs MSE is not discussed; a one-sentence justification would help, since the two losses emphasize different types of artifacts.","section":"Eq. (8)"},{"comment":"The footnote says iLRM predicts Gaussians at 2× downsampled resolution, resulting in 4× fewer Gaussians. This is correct for a uniform downsampling, but iLRM's 172K at 6 views is only modestly fewer than the dense 688K of DepthSplat. A clear explanation of why iLRM is not considered a 'dense' baseline would avoid confusion about the 5.7× claim.","section":"Table 2, footnote"},{"comment":"The notation N_init and N_final is clear, but the variable L_ATE is only defined in the text. Consider defining it in the equation or just below it to improve readability.","section":"Sec. A.3, Eq. (11)"}],"recommendation":"major_revision","confidential_remarks":"I agree with the stress-test concern: the paper's central claim is overstated. The efficiency contribution is real and well-supported by the ablations, but the 'state-of-the-art' phrasing in the abstract and conclusion is contradicted by Table 1, and the high-resolution baseline has a protocol mismatch. The requested changes are local (claim narrowing, baseline re-evaluation or explicit caveat) and should be addressable in a revision. I would not reject, but the current version is not acceptable as is."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Net take: the core contribution is real and the DL3DV numbers are strong, but the abstract overstates 'SOTA' and the high-res baseline is run off its training resolution. This is a solid paper that needs claim-narrowing, not a rework.\n\nWhat's genuinely new: ATSplat replaces pixel-aligned Gaussian regression with a sparse set of 3D anchor tokens initialized from coarse patch depth, decodes Gaussians as local offsets, and expands tokens with highest predicted uncertainty under a learned score supervised by D-SSIM error maps. The ablations make the case that each component matters: anchor+offset beats pixel-aligned and direct xyz; uncertainty selection beats random, FPS, and STE; and expansion monotonically improves quality. On DL3DV at 256x448, the method is clearly ahead: 27.28 PSNR with 120K Gaussians vs 24.19 for DepthSplat with 688K.\n\nSoft spots, in proportion. First, the abstract claims 'state-of-the-art rendering quality' but Table 1 on RealEstate10K shows iLRM beats it on PSNR and SSIM, and TokenGS ties on SSIM. The paper body correctly says 'comparable' — the abstract should match. Second, Table 3's high-res comparison uses a DepthSplat public checkpoint trained at 448x768, evaluated at 512x960; that's an unfair comparison. Against iLRM the margin is only 0.5 dB. Third, the 5.7x reduction is true when compared with dense per-pixel baselines, but not against iLRM or TokenGS, which already use fewer Gaussians; the claim is selective. Fourth, the uncertainty head is self-supervised by the model's own rendered error maps and validated only qualitatively (Fig. 4). The ablations show it works, so this isn't fatal, but a correlation/calibration plot and failure cases would make the central mechanism much easier to trust.\n\nWho this is for: anyone working on feed-forward 3DGS, efficient novel-view synthesis, or learned primitive allocation. The method is a worthwhile data point and the paper is clearly written.\n\nShould it be refereed? Yes. Send to peer review. The main fixes are to narrow the abstract's claims, replace or relabel the high-res DepthSplat comparison, and add a quantitative analysis of the uncertainty scores. No code is provided, but that's not a reason to desk-reject.","headline":"Core idea is solid and DL3DV results are strong, but the abstract overclaims SOTA and the high-res baseline is run off its training resolution.","tokens_in":19913,"tokens_out":3465,"would_cite":true,"duration_ms":29136,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ATSplat restores scene-adaptive Gaussian allocation in feed-forward 3DGS by starting from sparse 3D anchor tokens and expanding only the tokens predicted to be under-reconstructed, matching or beating dense pixel-aligned baselines with more","keywords":["3D Gaussian Splatting","feed-forward reconstruction","novel-view synthesis","adaptive capacity allocation","uncertainty-guided expansion","sparse 3D tokens","multi-view reconstruction","real-time rendering"],"falsifier":"Measure the rank correlation between the predicted uncertainty scores and the actual D-SSIM error maps on a held-out dataset with different scene statistics than training (e.g., cross-dataset evaluation). If the correlation is near zero or negative, the expansion targets are effectively arbitrary and the claimed compression gains would not survive when the selection ratios are replaced by oracle or random selection; alternatively, compare against an oracle that expands the true highest-error regions and check whether the quality gap closes.","tokens_in":18920,"feed_emoji":"📷","tokens_out":8738,"duration_ms":65255,"temperature":0.7,"pith_summary":"ATSplat argues that the dense pixel-aligned design shared by most feed-forward 3D Gaussian Splatting methods is wasteful: when one Gaussian is regressed per input pixel, the Gaussian budget and its spatial distribution are set by image resolution and camera sampling rather than by scene complexity. The paper proposes instead a sparse-to-adaptive pipeline: coarse patch-level depths are unprojected into sparse 3D anchor tokens, each token decodes into a cluster of Gaussians whose centers are learned offsets from the anchor, and an Adaptive Token Expansion (ATE) module predicts which tokens sit in poorly reconstructed regions and expands those tokens. On RealEstate10K and DL3DV, the authors report that this formulation matches or exceeds the rendering quality of dense feed-forward baselines while using more than 5.7x fewer Gaussians, reconstructing a high-resolution scene in under a second on one GPU and rendering at 1136 FPS. A sympathetic reader would care because it suggests feed-forward 3DGS quality is governed by where capacity is allocated, not by how densely primitives are sampled.","feed_headline":"ATSplat matches dense methods with 5.7x fewer Gaussians","feed_subtitle":"Sparse 3D anchors plus uncertainty-guided expansion put capacity where scenes are hard, not where pixels are packed.","key_machinery":"The two central components are the adaptive 3D anchor tokens and the Adaptive Token Expansion (ATE) module. Anchor tokens are obtained by unprojecting coarse patch features along predicted ray-depth and aggregating local 3D context; each token subsequently predicts K Gaussians whose centers are learned offsets from the anchor, so primitive placement is no longer constrained to input pixel grids. The ATE module is a lightweight MLP that outputs a per-token uncertainty score and is supervised by comparing a rasterized uncertainty map to the D-SSIM error map of intermediate renderings. At decoder blocks, the top fraction of tokens by predicted uncertainty are each replaced with M child tokens t","core_discovery":"The core discovery is that a feed-forward 3D Gaussian Splatting system can recover the scene-adaptive capacity allocation of optimization-based 3DGS. ATSplat does this by building a sparse 3D scaffold from coarse depth patches, decoding each anchor into K local Gaussians placed at learnable 3D offsets, and progressively expanding tokens whose predicted uncertainty exceeds a threshold, where the uncertainty head is trained to reproduce the model's own intermediate D-SSIM error maps. The reported result is that from 12 input images at 512x960, the model reconstructs a scene in less than a second into 311K Gaussians and renders novel views at 1136 FPS, matching or beating dense pixel-aligned fe","pith_inferences":["Editorial: If the uncertainty head stays calibrated across distribution shift, the same sparse-to-adaptive recipe could be applied to other primitive-based reconstruction models (e.g., large reconstruction models or voxel-based splatting) to reduce memory and bandwidth in on-device or streaming settings with minimal changes.","Editorial: A natural testbed is to make the selection policy explicitly budget-aware — e.g., train a controller to choose expansion ratios or thresholds per scene to hit a target Gaussian count — which would turn the current fixed-ratio mechanism into a true scene-dependent resource allocator.","Editorial: The paper's uncertainty signal could be reused as a heuristic for subsequent densification or pruning stages, effectively giving feed-forward methods the adaptive density-control loop that per-scene 3DGS enjoys, but in a single pass."],"forward_implications":["Feed-forward 3DGS can be compact without sacrificing rendering quality, since primitive count and placement become functions of scene complexity rather than input resolution or number of views.","A single forward pass on a commercial GPU can produce a high-resolution 3D representation in under a second and render it at interactive rates (1136 FPS), making on-the-fly reconstruction practical for real-time applications.","The uncertainty scores are a usable confidence map: regions that get expanded are exactly regions the model expects to render poorly, so the same signal can drive view selection, active capture, or downstream geometry tasks.","Because expansion happens inside the decoder, the additional tokens continue to cross-attend to input images, so allocated capacity comes with extra computation where it is needed, not just extra memory.","The authors' own dynamic-budget experiments show the fixed selection ratios can be replaced by uncertainty thresholds at inference, yielding scene-dependent Gaussian counts and pointing toward a fully adaptive compute allocator."],"fun_headline_variants":["ATSplat: 311K Gaussians, 1136 FPS, <1s build","Sparse anchors plus uncertainty expansion: 5.7x fewer Gaussians","Adaptive 3D tokens restore optimization-like capacity allocation","12 images to 311K Gaussians in under a second, 1136 FPS views","Feed-forward 3DGS that adapts primitive budget to scene complexity"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The learned per-token uncertainty score, supervised only by the model's own intermediate rendering errors during training, continues to rank the right tokens as needing expansion on unseen scenes at inference.","fun_headline_variants_meta":{"raw":{"variants":["ATSplat: 311K Gaussians, 1136 FPS, <1s build","Sparse anchors plus uncertainty expansion: 5.7x fewer Gaussians","Adaptive 3D tokens restore optimization-like capacity allocation","12 images to 311K Gaussians in under a second, 1136 FPS views","Feed-forward 3DGS that adapts primitive budget to scene complexity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001901,"raw_usage":{"total_tokens":7356,"prompt_tokens":883,"completion_tokens":6473,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":6369}},"tokens_in":627,"tokens_out":6473,"duration_ms":36942,"temperature":1.0,"reasoning_tokens":6369,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T09:51:58.594692+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the rank correlation between the predicted uncertainty scores and the actual D-SSIM error maps on a held-out dataset with different scene statistics than training (e.g., cross-dataset evaluation). If the correlation is near zero or negative, the expansion targets are effectively arbitrary and the claimed compression gains would not survive when the selection ratios are replaced by oracle or random selection; alternatively, compare against an oracle that expands the true highest-error regions and check whether the quality gap closes.","supporting_citations":[],"review_version":1}