{"id":"db9213de-3976-4f81-a7e7-421e1c2a4b90","arxiv_id":"2608.01186","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A dual-branch query decoder, guided by a frozen pretrained 3D vision model, decouples geometry from appearance to produce sharper pose-free feed-forward 3DGS novel views than prior baselines.","lead":"QuerySplat is a feed-forward method that turns a few unposed photos into a 3D Gaussian scene in one pass, using a frozen 3D-vision model for geometry and a separate learned branch for texture. On the DL3DV benchmark it reports the highest novel-view quality among compared pose-free and posed feed-forward 3DGS systems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Appendix E's loss-rank filter retaining 95% of training samples may inflate the reported DL3DV SOTA gains; its effect is unexamined.","rationale":"The reader identified the loss-rank filter as a secondary concern and the VGM geometry/depth/pose accuracy as the weakest premise. I view the loss-rank filter as the more load-bearing issue because it directly threatens the validity of the headline quantitative claim. The VGM-depth early-stage regularization ablation (Table 3) shows only a 0.29 dB drop when removed, so that specific component is not central to the SOTA gains. The VGM pose accuracy concern is diffuse: pose-free baselines in the comparison use the same VGM-style camera estimation protocol, and the architecture ablation (Table 2) already demonstrates the dual-branch decoder contributes ~2 dB over one-branch within the same VGM backbone. The loss-rank filter, however, is an unexamined training-time mechanism that selects the training distribution. Discarding the 5% highest-loss samples can increase average test metrics on an unfiltered test set if the test set's easy/hard ratio resembles the filtered training distribution, because the model never learns the hard tail. The paper reports no sensitivity analysis, no characterization of what is discarded, and no no-filter baseline. Figure 13's caption confirms the filter causes an abrupt loss decrease, showing it has a significant effect on optimization. This is a specific, testable, and plausible source of inflation for exactly the numbers that support the SOTA claim. Therefore, the central claim should be provisionally accepted only if this experiment is run; the reader's CONDITIONAL verdict remains appropriate. My agreement is partial because the reader's weakest_assumption prioritized VGM geometry/depth, whereas I rank the loss-rank filter as the more direct threat to the headline empirical claim.","tokens_in":20161,"tokens_out":10062,"duration_ms":99467,"concrete_test":"Retrain QuerySplat under the identical protocol (base stage + progressive finetuning, 64 A800 GPUs, same hyperparameters) but with the loss-rank filter disabled (100% retention). Keep all other settings unchanged. Evaluate on the same DL3DV-Evaluation interpolation splits (2/4/12 views; large/medium/small) and compare average PSNR/SSIM/LPIPS against the reported values (e.g., 4-view interpolation PSNR 24.5765). If the differences are within 0.1 dB, the concern is resolved. If the 4-view PSNR drops by more than 0.5 dB, or the 2.30 dB and 1.04 dB gains over the best pose-free and pose-required baselines shrink substantially, the filter is a key contributor and the headline claim must be qualified. Additionally, run a retention-rate sweep at 90%, 95%, and 98% to map sensitivity, if compute permits.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that QuerySplat achieves SOTA pose-free feed-forward 3DGS, with average PSNR gains of 2.30 dB over the best pose-free baseline and 1.04 dB over the best pose-required baseline. Appendix E states: 'During late training, a loss-rank filtering strategy retains 95% of samples according to their historical reconstruction losses to reduce the influence of unstable training cases.' This is a sample-selection mechanism that discards the 5% of training samples with the highest historical reconstruction loss. Such samples are likely those where VGM-predicted geometry, depth, or poses are poor, since rendering loss correlates with geometric failure. By training on the remaining easy 95%, the model may overfit to an easy subset of the distribution, and because the evaluation set is not filtered, the reported average PSNR could be inflated relative to a model trained on all data. The paper provides no ablation of the retention rate, no characterization of the discarded samples, and no comparison to a no-filter baseline. Figure 13's caption explicitly notes that an abrupt decrease in the loss curve results from this filtering strategy, indicating it materially alters training dynamics. Since the headline SOTA numbers are averages over DL3DV-Evaluation, this unexamined training-time selection mechanism is load-bearing: the reported gains might partly come from avoiding hard cases rather than from the decoupled architecture itself.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"QuerySplat proposes a feed-forward 3D Gaussian Splatting predictor built around a dual-branch query decoder. Geometry queries attend to features from a frozen Vision Geometric Model (VGGT-Omega) and predict positions, scales, and rotations; appearance queries attend to RGB and Plücker ray embeddings and predict opacities and spherical-harmonic colors. The VGM also supplies camera poses and a Sim(3)-aligned coordinate system, which the paper argues makes the method pose-free. A Chamfer-distance and opacity-floor regularizer is applied only in early training to stabilize initialization. The model is trained on DL3DV and evaluated on DL3DV-Evaluation, reporting state-of-the-art interpolation results: average PSNR gains of 2.30 dB over the best pose-free baseline and 1.04 dB over the best pose-required baseline, with optional test-time optimization providing further gains.","tokens_in":20448,"tokens_out":6035,"duration_ms":61855,"significance":"If the claims hold, this is a meaningful advance: it shows that decoupling geometry and appearance in query-based 3DGS, combined with a frozen VGM prior, can match or exceed pixel-aligned feed-forward methods while remaining pose-free and fast. The controlled ablations in Table 2 (one-branch fusion drops PSNR by 1.98 dB, removing appearance by 4.25 dB) and the backbone swap in Table 4 (VGGT vs. VGGT-Omega within 0.18 dB) are genuine strengths and provide evidence that the architectural separation, not just the VGM, matters. However, the headline SOTA claim is weakened by an unexamined training-time sample filter that removes the 5% highest-loss training samples, and by the fact that all architectural ablations are run at a much smaller scale than the final model. These issues need to be resolved before the quantitative claims can be considered established.","major_comments":[{"comment":"Appendix E states: 'During late training, a loss-rank filtering strategy retains 95% of samples according to their historical reconstruction losses.' This discards the 5% of training samples with the highest historical loss, i.e., the hardest cases. Since evaluation on DL3DV-Evaluation is not filtered, the model may be trained on an easy subset while tested on the full distribution, potentially inflating the headline average PSNR gains in Table 1. Figure 13's caption explicitly attributes an abrupt loss decrease to this filter, showing it materially alters training dynamics. There is no ablation of the retention rate, no characterization of discarded samples, and no comparison to training on all data. Because the central SOTA claim (2.30 dB over best pose-free, 1.04 dB over best pose-required) is an unfiltered evaluation average, this is a load-bearing unexamined choice; please add a no-","section":"Appendix E / Figure 13"},{"comment":"All architecture ablations (Tables 2-4) are trained only for 150K base iterations with 1,024 queries and no progressive query scaling, while the final model uses a 300K base stage plus progressive finetuning to 8,192 queries and the loss-rank filter. The paper argues the relative ordering is stable, but the magnitudes of the architecture benefits (e.g., 1.98 dB and 4.25 dB in Table 2) are not shown to persist in the final configuration, and the effect of progressive query scaling is not isolated. Since the headline comparison is the full pipeline, please demonstrate that the decoupling benefit remains at or near the final training protocol, at minimum without the loss-rank filter.","section":"Ablation Studies / Appendix D"},{"comment":"The pose-free property and the early-stage regularization both rest on the accuracy and consistency of the frozen VGM. The Sim(3) alignment between the input-only and all-view VGM passes is assumed drift-free, and the Chamfer target in Eq. (6) is a point cloud back-projected from VGM depth. No pose or depth error analysis is reported (e.g., rotation/translation error against COLMAP on DL3DV-Evaluation), and no robustness test is provided. If VGM poses or depths are biased on some scene types, the coordinate alignment and the cold-start regularization could misdirect training. Please add a quantitative evaluation of VGM pose/depth accuracy and a sensitivity study with perturbed poses or depths.","section":"Self-Calibrated Coordinate System / Eq. (6)"}],"minor_comments":[{"comment":"The parenthetical '(Note that each global step contains 500 (0.5k) steps with totally 150k steps)' is confusing and should be clarified or removed.","section":"Figure 13 caption"},{"comment":"The in-the-wild TokenGS comparison concedes that 'scene sizes are not strictly consistent' due to TokenGS's rectangular input requirement. Please state precisely how this affects the qualitative comparison and whether the shown results are cropped or resized differently.","section":"Appendix B"},{"comment":"The paper reports averages over 300 cases per split without error bars or confidence intervals. Given the headline SOTA claim, please report standard errors or significance tests, especially for the 1.04 dB pose-required margin.","section":"Tables 5-13"},{"comment":"The limitations paragraph mentions the fixed query budget but does not quantify its effect on very large or complex scenes. A brief quantitative assessment (e.g., PSNR vs. scene extent) would be useful.","section":"Appendix A"},{"comment":"The VGM backbone is referred to as 'VGGT-Ω' in the text and 'VGGT-Omega' in some places; normalize the notation.","section":"Notation"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the core architectural idea is plausible, with informative controlled ablations. The main blocker is the unexamined loss-rank filter in Appendix E, which directly affects the interpretation of the headline SOTA numbers. I am not concerned about circularity: the model is trained with photometric rendering loss on held-out evaluation cases and compared against external baselines. I would also ask the editor to ensure that the dataset-level calibration coefficient used in the supplementary TokenGS comparison does not leak into the main evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuine architectural improvement over TokenGS and the ablations support the decoupling story. The headline DL3DV gains are plausible but not independently checkable until code or weights ship, and the Appendix E loss-rank filter is a real unexamined confound.\n\nWhat's new: the dual-branch query decoder that gives geometry queries and appearance queries separate feature/query pathways, plus the early-stage Chamfer/opacity-floor regularization guided by frozen VGM depth. The paper is careful to test the core hypothesis: one-branch fusion drops PSNR by ~2 dB and removing the appearance branch by ~4 dB, and the VGGT/VGGT-Omega swap stays within 0.18 dB. That backbone swap is a nice piece of evidence that the mechanism, not the particular encoder, does the work. They also disclose the non-like-for-like TokenGS comparison in the appendix, which is honest.\n\nSoft spots. The Appendix E loss-rank filter retaining 95% of samples is genuinely load-bearing and unexamined. The training loss curves in Figure 13 show an abrupt drop at the filter's introduction, and no ablation varies the retention rate or compares against no filter. The discarded samples are probably exactly the VGM-failure cases, so the reported averages could be inflated relative to training on all data. This is a moderate concern, not a fatal one—the filter is shared among ablations, so it doesn't explain the relative improvements, but it could exaggerate the absolute SOTA numbers. Second, there are no error bars anywhere and no released code or weights; with a 64-GPU training setup that's understandable but it means the headline numbers rest on a single run. Third, the ablations run only 150K base iterations without progressive query scaling; the authors argue convergence order is stable, and the loss curves support that, but it's still a shortened protocol. Minor: the Sim(3) alignment between two VGM passes is assumed drift-free; that's worth a sentence in a revision.\n\nBottom line: the central architectural claim holds up on the evidence in the paper. The empirical magnitudes deserve independent confirmation. I'd send this to peer review—the contribution is real and the paper is readable—but I'd demand a no-filter baseline, variance or error bars, and a code/weights release before trusting the exact dB numbers.","headline":"Strong engineering contribution with an honest ablation story; the Appendix E sample filter and missing code/variance are the real reasons to hold before trusting the exact SOTA numbers.","tokens_in":21023,"tokens_out":1879,"would_cite":true,"duration_ms":18506,"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":"QuerySplat separates geometry from appearance in a query decoder and reaches top DL3DV novel-view PSNR without input poses.","keywords":["3D Gaussian Splatting","feed-forward reconstruction","novel view synthesis","pose-free reconstruction","query-based decoder","geometry-appearance decoupling","vision geometric model","DL3DV"],"falsifier":"Compare QuerySplat against the same model fed oracle ground-truth poses on DL3DV: if the PSNR/SSIM gap is large, VGM camera error, not the decoder design, drives the reported gains. Separately, disable the early Chamfer/opacity regularizers on a set of scenes where VGM depth is known to be biased (mirrors, reflective floors, low-texture walls): if quality does not degrade, the depth-guided cold start is not the operative mechanism. Either observation would distinguish the decoupling hypothesis from the prior-quality hypothesis.","tokens_in":19981,"feed_emoji":"🎥","tokens_out":8608,"duration_ms":81277,"temperature":0.7,"pith_summary":"QuerySplat claims that the blurriness and geometric disorder of earlier query-based feed-forward 3D Gaussian Splatting come from forcing one set of queries to predict both spatial and appearance attributes at once. It builds a dual-branch decoder: a frozen pretrained vision geometric model supplies geometry features, cameras, and depth for a geometry branch, while a separate appearance branch attends to image and ray features to predict color and opacity. On DL3DV, the paper reports average PSNR gains of 2.30 dB over the best pose-free baseline and 1.04 dB over the best pose-required baseline, with sharper textures and more coherent Gaussian structure. A careful reader would care because it suggests sparse-view, pose-free 3D reconstruction from casual photos can be both fast and high-fidelity, without needing per-scene optimization or external camera calibration.","feed_headline":"Split geometry and appearance lifts pose-free novel views by 2.3 dB","feed_subtitle":"QuerySplat's separated geometry and appearance queries beat posed baselines with no external cameras.","key_machinery":"The key mechanism is the attribute-aware dual-query decoder: two transformer decoder stacks with separate learnable queries. Geometry features come from a frozen VGGT-Ω (fused from several intermediate layers); appearance features combine RGB patch embeddings with Plücker ray embeddings computed from VGM-predicted cameras. The geometry branch emits positions, scales, and rotations as scene-level slots; the appearance branch reuses those geometry tokens as base state and adds queries for opacity and spherical-harmonic color. The procedure is completed by self-calibrated coordinate frames—two VGM passes aligned by Sim(3)—and transient early-stage regularization (bidirectional Chamfer distance","core_discovery":"The central claim is that Gaussian attributes have different information needs, and a query-based 3DGS predictor should decouple them. Geometry queries attend only to VGM geometry features and decode centers, scales, and rotations; appearance queries start from the decoded geometry tokens, add learnable queries, attend to RGB patch embeddings and Plücker rays, and decode opacities and spherical-harmonic colors. The VGM is frozen and defines a self-calibrated coordinate system, with an input-only pass for reconstruction and an all-view pass for supervision cameras aligned by Sim(3); an early Chamfer-distance and opacity-floor regularizer pulls the first Gaussian configurations toward the VGM","pith_inferences":["Beyond the paper's DL3DV experiments, the same dual-branch logic could be tested on other feed-forward representations (triplanes, voxels, mesh generators) where geometry and appearance also have different frequency and locality demands.","The paper's own Appendix A limits the fixed query budget for very large scenes; a reader might infer that region-splitting and merging is the next scaling step, but it is not implemented here.","Because early regularization leans on VGM depth, one can predict that scene classes where VGM depth is systematically biased (mirrors, transparent objects, repetitive texture) will cluster the failures; the paper does not report per-class breakdowns.","The pose-free comparison is clean only among native pose-free methods; as the appendix concedes, adapting a pose-required baseline like TokenGS with external VGGT-Ω cameras is not like-for-like, so the 1.04 dB posed-baseline gap should be read with that caveat."],"forward_implications":["If correct, query-based non-pixel-aligned 3DGS is no longer inherently blurry: appearance-specific queries recover high-frequency detail while geometry queries keep scene coherence.","Pose-free reconstruction becomes a default property of the framework rather than an added module, since VGM-predicted cameras and coordinates replace external pose input.","The decoupling recipe transfers across VGM backbones (VGGT vs VGGT-Ω give comparable results), suggesting geometric-prior consumption, not a specific encoder, drives the benefit.","Optional test-time optimization in feature space provides an accuracy/runtime trade-off that improves even extrapolation views, not just input-view memorization.","Applications like repairing a 3D scene from video or reconstructing from text-to-video frames become feasible because the same feed-forward model lifts restored or generated 2D frames back into coherent 3DGS."],"supporting_citations":[{"why":"Frozen VGGT-Ω encoder that supplies geometry-aware features, camera poses, intrinsics, depth, and the self-calibrated coordinate frame.","marker":"Wang et al. 2026a"},{"why":"Defines 3D Gaussian Splatting, the differentiable rendering representation and supervision target that QuerySplat predicts.","marker":"Kerbl et al. 2023"},{"why":"TokenGS, the query-based non-pixel-aligned 3DGS baseline whose entanglement and blurriness QuerySplat targets, and source of the visibility-loss idea.","marker":"Ren et al. 2026"},{"why":"DL3DV dataset used for training and the large-scale evaluation benchmark where the headline PSNR gains are measured.","marker":"Ling et al. 2024"},{"why":"VGGT, the alternative geometric backbone in the backbone-ablation experiment that shows the decoder is not tied to one VGM.","marker":"Wang et al. 2025a"},{"why":"AnySplat, a pose-free pixel-aligned baseline that QuerySplat compares against on DL3DV.","marker":"Jiang et al. 2025"},{"why":"YoNoSplat, a pose-free feed-forward 3DGS baseline used in the quantitative comparisons.","marker":"Ye et al. 2025a"}],"fun_headline_variants":["Geometry and appearance split sharpens pose-free views","QuerySplat decouples shape and look for 2.3 dB gain","Pose-free 3D scene builder beats posed ones by 2.3 dB","Splitting queries by role lifts novel view synthesis","No camera poses needed - QuerySplat renders sharper"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that the frozen VGM's predicted cameras, geometry features, and depth are accurate and consistent enough to define the coordinate system and the early training anchor; if its depth is biased on some scene type, the Chamfer cold-start could misdirect the Gaussian centers, and the pose-free claim inherits whatever camera error the VGM has.","fun_headline_variants_meta":{"raw":{"variants":["Geometry and appearance split sharpens pose-free views","QuerySplat decouples shape and look for 2.3 dB gain","Pose-free 3D scene builder beats posed ones by 2.3 dB","Splitting queries by role lifts novel view synthesis","No camera poses needed - QuerySplat renders sharper"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000824,"raw_usage":{"total_tokens":3446,"prompt_tokens":752,"completion_tokens":2694,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":2606}},"tokens_in":496,"tokens_out":2694,"duration_ms":20152,"temperature":1.0,"reasoning_tokens":2606,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T00:27:01.928786+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare QuerySplat against the same model fed oracle ground-truth poses on DL3DV: if the PSNR/SSIM gap is large, VGM camera error, not the decoder design, drives the reported gains. Separately, disable the early Chamfer/opacity regularizers on a set of scenes where VGM depth is known to be biased (mirrors, reflective floors, low-texture walls): if quality does not degrade, the depth-guided cold start is not the operative mechanism. Either observation would distinguish the decoupling hypothesis from the prior-quality hypothesis.","supporting_citations":[],"review_version":1}