{"id":"589a00e0-99b8-4198-9c6c-734d17061442","arxiv_id":"2608.07937","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FlexSplat jointly trains a geometry estimator with a query-based Gaussian decoder, reconstructing objects from uncalibrated photos within 0.7 dB PSNR of posed baselines on GSO.","lead":"FlexSplat rebuilds a 3D object from a few photos without being told where the camera was, using a pretrained geometry network that is fine-tuned together with a set of 3D Gaussian shapes. Its quality on object benchmarks is close to methods that do receive camera positions, which could make 3D asset creation and robotic manipulation simpler.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Predicted-camera translation error is the untested mode that could break the cross-view consensus; without a translational analogue of Table 7, the calibration-free parity claim rests on an assumption the paper explicitly defers.","rationale":"The reader's weakest_assumption is the cross-view consistency of predicted geometry; my independent read converges on the same point and sharpens it: the missing translational perturbation test is the exact experiment that would validate or refute the claim. The paper has genuine strengths: the joint fine-tuning of VGGT is a real design choice, ablations (Table 6) isolate its contribution, and the code is promised; those support a conditional acceptance rather than rejection. But the central quantitative claim (within 0.69 dB of UniGS without poses) is only as strong as the robustness of the predicted cameras. Sec. 4.9 tests rotation only and the limitations explicitly defer translation. Eq. 2's per-view center definition and the undeclared fusion into a single world-space µ is a second, related gap: the text assumes what the mechanism must enforce. A concrete translational perturbation study plus a code-level check of the center fusion would settle whether the concern lands. If translation sensitivity turns out to be large, the headline comparison would need to be qualified; until then, CONDITIONAL remains the appropriate verdict and no change to the reader's assessment is needed. One further minor inconsistency worth correcting: Sec. 3.2 says the target camera π_j' is 'predicted, not given,' but the evaluation in Sec. 4.2 uses a global alignment so that each held-out target view can be rendered and compared against its ground truth, implying the target camera is supplied for evaluation; this wording should be clarified but is not the primary load-bearing issue.","tokens_in":12900,"tokens_out":8531,"duration_ms":97379,"concrete_test":"Add a translational analogue of Table 7: on the 250 GSO objects, apply independent random translations to the estimated camera centers with magnitudes 1%, 3%, 5% of object scale (rotations fixed), and recompute PSNR/SSIM/LPIPS with the same evaluation protocol. If the drop at 3% (the geometry head's own error level) exceeds the 0.69 dB margin to UniGS, the cross-view consensus claim fails for the actual operating error. Also report per-view relative translation error after global similarity alignment. In parallel, inspect the released code to determine how Eq. 2's per-view centers are combined into the µ_n projected in Eq. 3; if there is no explicit fusion, re-derive Eq. 3 from a single world-space center or state which view anchors the parameterization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"FlexSplat's central claim (Sec. 4.4: within 0.69 dB PSNR of posed UniGS with no poses) rests on the decoder's ability to absorb errors in the jointly predicted geometry. The mechanism is cross-view consensus in Sec. 3.3: Eq. 3 computes reference points by projecting the same world-space Gaussian center µ into each view using predicted R_i, t_i, and Eq. 2 grounds the center in predicted depth. This is only sound if the predicted cameras live in one shared coordinate frame and the per-view projections of µ correspond to the same surface point. The paper evidences robustness only for rotation: Table 7 adds 1-20° rotational noise and explicitly defers translational sensitivity to future work (Sec. 4.9, Sec. 5). But the evaluation alignment (Sec. 4.2) is only a global similarity transform; it can correct common scale/rotation/translation, not per-view relative translation inconsistencies. The geometry head's own operating point has 3% object-scale camera-center error (Sec. 4.9), and no experiment shows how much of that error is relative. If relative translation errors of this magnitude shift the projections in Eq. 3 and the depth anchors in Eq. 2, the claimed 0.69 dB gap could be an artifact of favorable global alignment rather than of genuine cross-view consensus. Relatedly, Eq. 2 defines the Gaussian center only 'in the camera frame of view i'; the paper never states how the per-view centers µ_{n,i} are fused into the single world-space µ_n used in Eq. 3, so shared-frame consistency is assumed rather than derived.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FlexSplat, a feed-forward query-based 3D Gaussian splatting method for novel view synthesis from uncalibrated, object-centric multi-view images. It replaces the posed-input assumption of prior query-based methods (LeanGaussian, UniGS) with a jointly fine-tuned VGGT geometry transformer that predicts per-image cameras and depth. These predictions are used in three ways: a depth-guided Gaussian mean parameterization (Eq. 2), multi-view deformable cross-attention over fused depth/appearance features (Eq. 3), and an uncertainty-weighted depth-consistency loss (Eq. 5). Experiments on ShapeNet-SRN and GSO report single-view and four-view results, ablations, inference times, and rotation-perturbation robustness. The central claim is that FlexSplat matches or approaches posed state-of-the-art reconstructors, staying within 0.69 dB PSNR of UniGS on four-view GSO while using no camera poses or ground-truth depth.","tokens_in":13243,"tokens_out":7411,"duration_ms":79643,"significance":"If fully supported, the result would be significant: it would demonstrate that a jointly trained geometry front-end can make query-based Gaussian reconstruction calibration-free at a small fidelity cost, while keeping a compact representation whose primitive count is independent of image resolution. The paper has concrete strengths: the ablation in Table 6 isolates a 1.38 dB benefit from joint training of VGGT, and the rotation-perturbation study in Table 7 gives initial evidence for the cross-view consensus mechanism. However, the stress-test concern about per-view translation errors does land. The paper never specifies how the per-view Gaussian centers of Eq. (2) are fused into the world-space center used in Eq. (3), the evaluation protocol for target-view poses is underspecified, and the robustness study explicitly covers only rotation noise. These omissions leave the headline parity claim less supported than the abstract suggests. No error bars or multiple seeds are reported, so the 0.69 dB gap to UniGS cannot be assessed statistically.","major_comments":[{"comment":"The paper never states how the per-view Gaussian centers mu_{n,i}, defined in the camera frame of view i by Eq. (2), are fused into the single world-space center mu_n that Eq. (3) projects into every view. This is not a presentational detail: the cross-view consensus mechanism in Sec. 3.3 is only sound if the predicted cameras share a common coordinate frame and the projections of mu_n correspond to the same surface point. Please specify the fusion operation explicitly (e.g., inverse transformation to world coordinates followed by averaging, or a learned combination) and provide evidence that the jointly predicted cameras are mutually consistent beyond a global similarity. The paper's own robustness study in Table 7 perturbs only rotation and Sec. 4.9 defers translational sensitivity to future work; because the evaluation alignment in Sec. 4.2 is only a global similarity transform, it cannot correct per-view relative translation errors. Given the reported 3% object-scale camera-center error, a translational analogue of Table 7 is needed to support the claim that the decoder absorbs residual pose error.","section":"Sec. 3.2-3.3, Eq. (2)-(3)"},{"comment":"The protocol for target-view poses is underspecified. Sec. 3.2 states that the target-view camera pi_j' is predicted, not given, while Sec. 4.2 describes only a global similarity alignment of the estimated cameras for evaluation. It is unclear whether held-out target views are rendered at predicted cameras and then compared with the ground-truth image at the corresponding dataset viewpoint, or at ground-truth target cameras. If target poses are predicted, the RGB supervision in Eq. (5) and the evaluation both depend on the same predicted geometry, which could bias the results; if ground-truth target poses are used for rendering while only input poses are removed, the 'no poses' claim should be qualified accordingly. Please specify the exact training and evaluation rendering protocol, including how the similarity alignment is computed for target views.","section":"Sec. 3.2 and Sec. 4.2, Eq. (1)"},{"comment":"The headline comparison in Table 2 has no error bars or multiple seeds. The paper reports FlexSplat within 0.69 dB PSNR and 0.012 SSIM of UniGS and describes the LPIPS difference (0.041 vs 0.042) as 'within noise', but no variance estimates are provided anywhere in the paper. Without repeated runs or confidence intervals, the central claim that calibration-free operation incurs only a small fidelity cost is not statistically supported. Please report error bars over at least a few seeds, per-object standard errors, or another accepted measure of variability for the main tables and state the number of runs.","section":"Table 2 and Sec. 4.4"},{"comment":"The closely related pose-free query-based method C3G is cited in Sec. 2 and Sec. 3.4 as reconstructing scenes from unposed images with only about 2K Gaussians, but it is not included in Table 2 or Table 3. Since C3G is the most direct existing baseline for the paper's central claim of calibration-free query-based Gaussian reconstruction, its omission makes the comparison incomplete. Please add a comparison on GSO, or provide a clear quantitative justification for why C3G cannot be evaluated in this object-centric setting.","section":"Sec. 2, Sec. 3.4, and Table 2"}],"minor_comments":[{"comment":"Eq. (2) uses calibrated normalized image coordinates (u1i, u2i) but never defines how these are obtained for each Gaussian query; the phrase 'the corresponding location in view i' is ambiguous. Please define the reference location used for each query before back-projection.","section":"Eq. (2)"},{"comment":"The depth-consistency term in Eq. (5) aligns rendered depth with VGGT's own predicted depth, so it is a self-consistency regularizer rather than an external geometric signal. The text should state this explicitly to avoid any impression that ground-truth depth is used.","section":"Eq. (5)"},{"comment":"The claim that the Gaussian budget is 'not dictated by the number of views' is weakened by the fact that the experiments scale the budget with the view count (10K/15K/20K for 1/2/4 views). Please either hold the budget fixed in at least one experiment or soften the wording.","section":"Sec. 3.4 and Table 4"},{"comment":"The single-view GSO results show FlexSplat matching UniGS in LPIPS (0.149) but trailing in PSNR by 0.42 dB; the abstract's phrase 'matches the best perceptual (LPIPS) quality' should clarify that this refers to LPIPS only, not overall quality, and that the four-view and single-view claims are distinct.","section":"Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the core idea is promising. My main concern is that the calibration-free parity claim rests on a cross-view consistency assumption that is not stated precisely and is tested only for rotation. The authors should be required to specify the center-fusion operation, clarify the target-pose evaluation protocol, and add a translational perturbation experiment before the paper can be accepted. I did not find evidence of citation manipulation or a scope mismatch; the omission of C3G from the comparison is a substantive gap but one that can be addressed in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: FlexSplat is a real engineering advance, not a revolution. It shows that jointly fine-tuning a geometry transformer with a query-based Gaussian decoder removes the pose requirement for object-level NVS at a small fidelity cost. The main numbers support that claim; the two soft spots worth naming are the missing C3G comparison and the untested translational failure mode.\n\nWhat is genuinely new: the specific system, jointly trained VGGT with a depth-guided UniGS-style decoder, is a sensible combination and the paper does a good job isolating its value. The ablation showing that freezing VGGT costs 1.38 dB PSNR is the cleanest evidence that the joint training is load-bearing. Four-view GSO results (29.73 PSNR vs. UniGS's 30.42, with no poses) are credible, and the single-view ShapeNet table is on par with LeanGaussian. Code is public, which matters.\n\nNow the soft spots, in rough order of importance. First, the paper's own Related Work credits C3G with reconstructing from unposed images using roughly 2K Gaussians. FlexSplat does not compare with C3G or say what new capability it adds beyond that. That is a real novelty gap that a referee needs the authors to close. It may be that C3G operates under different assumptions, but as written the reader cannot tell.\n\nSecond, the cross-view consensus mechanism assumes the predicted cameras live in one shared coordinate frame and that per-view projections of a Gaussian center correspond to the same surface point. The paper does not define how the per-view centers from Eq. 2 are fused into the world-space mu used in Eq. 3. The rotation-only perturbation study is helpful, but the geometry head itself has 3% object-scale camera-center error, and no translational perturbation is reported. The evaluation's global similarity alignment can correct consistent global errors, not per-view relative translation inconsistencies. The authors explicitly defer translational sensitivity to future work, so this is a known limitation rather than a hidden flaw, but it leaves the headline parity claim partly unproven.\n\nThird, no error bars or multiple seeds. For a system trained on a large synthetic-ish mixture and evaluated on 250 GSO objects, one run is thin. This is a minor issue compared to the first two.\n\nBottom line: this paper is worth serious peer review. The contribution is useful for object-centric reconstruction and the evidence is largely consistent. A revision with a C3G comparison, a per-view translation perturbation study, and repeated evaluations would substantially strengthen it. I would take it to a reading group and would cite it, but I would not treat the calibration-free parity claim as settled yet.","headline":"A credible engineering contribution showing joint geometry training can make query-based Gaussian reconstruction calibration-free at small fidelity cost, but it needs a C3G comparison and a translational robustness study before the parity claim is fully earned.","tokens_in":13792,"tokens_out":1759,"would_cite":true,"duration_ms":23608,"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":"FlexSplat shows that jointly training a geometry transformer with a query-based Gaussian decoder removes the need for camera poses in feed-forward 3D reconstruction, matching posed state-of-the-art within 0.7 dB PSNR on object-centric…","keywords":["novel view synthesis","3D Gaussian splatting","feed-forward reconstruction","calibration-free","query-based transformer","multi-view deformable attention","depth-guided parameterization","object-centric reconstruction"],"falsifier":"Add translational noise of magnitude 3 percent of object scale—the head's own reported camera-center error—to the predicted cameras on the four-view GSO evaluation and measure the PSNR drop; if the drop exceeds the 0.69 dB gap to posed UniGS, then the cross-view consensus does not actually absorb the geometry error it claims to tolerate, since the paper currently probes only rotation perturbations.","tokens_in":12700,"feed_emoji":"🧊","tokens_out":8005,"duration_ms":79778,"temperature":0.7,"pith_summary":"FlexSplat asks whether feed-forward novel-view synthesis from a few photos of an object still works when camera calibration is not provided. The paper's answer is yes: a geometry transformer is trained jointly with a query-based Gaussian decoder, so the predicted per-image cameras and depth ground where each Gaussian sits and how it attends to the input views. On ShapeNet-SRN and Google Scanned Objects, this uncalibrated pipeline matches posed state-of-the-art reconstructors within 0.7 dB PSNR and matches the best perceptual (LPIPS) quality on GSO, while never seeing camera poses or ground-truth depth. The result matters because it removes the structure-from-motion preprocessing that feed-forward reconstruction was supposed to eliminate, and it keeps a compact Gaussian budget independent of image resolution and view count.","feed_headline":"No camera poses: FlexSplat rivals calibrated 3D reconstruction","feed_subtitle":"Jointly predicting cameras and depth keeps query-based Gaussians within 0.7 dB of posed methods on GSO.","key_machinery":"The central mechanism is the query-based Gaussian decoder grounded in predicted geometry. A set of learnable Gaussian queries, initialized near the estimated scene center (the least-squares intersection of predicted camera optical axes), is refined over three layers; at each layer, the center of each Gaussian is projected into every input view to produce reference points, around which multi-view deformable attention samples fused depth-and-appearance features and updates the query in world space. Because attention weights are normalized jointly over all views and sampling points, moderate per-view camera and depth error is averaged down rather than propagated—the cross-view consensus that the robustness experiments probe. The depth-guided mean parameterization (back-projecting predicted depth plus a learnable offset) stabilizes convergence, and the uncertainty-weighted depth-consistency objective in the loss couples the geometry transformer's training to the reconstruction task.","core_discovery":"The paper claims that a jointly trained geometry front-end is sufficient to bring calibration-free operation to query-based Gaussian reconstruction. FlexSplat inherits the unitary-Gaussian decoder of LeanGaussian and UniGS—learnable 3D ellipsoids refined by multi-view deformable cross-attention—but replaces the posed-input assumption with a VGGT-based transformer that predicts per-image cameras and dense depth, fine-tuned together with the decoder under rendering and depth-consistency objectives. Predicted depth seeds a depth-guided parameterization of each Gaussian center, fused depth-and-appearance features serve as keys and values for the cross-attention, and an uncertainty-weighted depth-consistency loss aligns rendered depth to the predicted depth. Because each Gaussian aggregates evidence from all views and the update is reconciled in world space, the decoder forms a cross-view consensus that tolerates the imperfect cameras and depth it is given. On four-view GSO, FlexSplat reaches 29.73 dB PSNR and 0.949 SSIM against 30.42 dB and 0.961 for posed UniGS, with best-in-comparison LPIPS of 0.041, indicating that calibration-free operation costs less than 0.7 dB while matching perceptual quality.","pith_inferences":["A natural extension the paper leaves open is scene-level pose-free reconstruction; whether the cross-view consensus transfers to background-heavy scenes is untested, and the paper's own limitation section flags this.","Since the paper's robustness study perturbs only rotation and explicitly leaves translational sensitivity to future work, the tolerance claim is only half-established; a systematic translation error in the predicted cameras might violate the shared-frame assumption more severely.","The 1.38 dB penalty for freezing VGGT implies that the benefit comes from co-adaptation with the decoder, not from the pretrained geometry alone; a testable prediction is that any strong geometry prior fine-tuned jointly would show a similar gain."],"forward_implications":["Camera calibration can be dropped from the input of query-based Gaussian reconstruction on object-centric data: the jointly trained geometry head supplies cameras and depth, and the decoder's cross-view consensus absorbs their residual error.","The Gaussian budget is a free hyper-parameter decoupled from image resolution and view count, so the representation stays compact—around $2\\times10^4$ primitives for four $256\\times256$ views versus about $2.6\\times10^5$ for pixel-aligned methods.","Because the geometry transformer is fine-tuned with the rendering and depth-consistency losses, its predicted depth and cameras are shaped by reconstruction quality rather than by standalone geometric accuracy, allowing partial self-correction.","The robustness result—only a 0.22 dB drop at 1 degree of rotation noise—suggests that moderate camera-estimation error is tolerated gracefully at inference, which would allow sloppy capture protocols without pose refinement."],"supporting_citations":[{"why":"Supplies the query-based Gaussian decoder with deformable attention that FlexSplat extends from the single-view posed setting to uncalibrated multi-view.","marker":"[29]"},{"why":"Defines the posed multi-view unitary-Gaussian baseline with multi-view deformable cross-attention; FlexSplat inherits its decoder and compares directly against it on GSO.","marker":"[30]"},{"why":"The geometry transformer used as the front-end, predicting per-image cameras and depth and jointly fine-tuned in FlexSplat.","marker":"[26]"},{"why":"The 3D Gaussian splatting representation and differentiable renderer underlying the whole pipeline.","marker":"[12]"},{"why":"Source of the depth-guided mean parameterization (Eq. 2) and a pixel-aligned baseline whose primitive count grows with resolution.","marker":"[23]"},{"why":"Provides the standard global-similarity alignment used to evaluate pose-free novel-view synthesis, allowing comparison against posed methods.","marker":"[35]"}],"fun_headline_variants":["FlexSplat: no poses needed, matches calibrated 3D","Calibration-free Gaussians rival posed methods","Joint camera-depth prediction closes gap to posed NVS","FlexSplat: 0.7 dB from posed, best LPIPS on GSO","No point clouds, no poses: FlexSplat reconstructs views"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The cameras and depths predicted by the jointly trained geometry head are consistent across views in a single shared coordinate frame, so that projecting a Gaussian's center into every view samples the same surface point in each image.","fun_headline_variants_meta":{"raw":{"variants":["FlexSplat: no poses needed, matches calibrated 3D","Calibration-free Gaussians rival posed methods","Joint camera-depth prediction closes gap to posed NVS","FlexSplat: 0.7 dB from posed, best LPIPS on GSO","No point clouds, no poses: FlexSplat reconstructs views"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00023,"raw_usage":{"total_tokens":1555,"prompt_tokens":1094,"completion_tokens":461,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":710,"completion_tokens_details":{"reasoning_tokens":373}},"tokens_in":710,"tokens_out":461,"duration_ms":5558,"temperature":1.0,"reasoning_tokens":373,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:38:59.189256+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Add translational noise of magnitude 3 percent of object scale—the head's own reported camera-center error—to the predicted cameras on the four-view GSO evaluation and measure the PSNR drop; if the drop exceeds the 0.69 dB gap to posed UniGS, then the cross-view consensus does not actually absorb the geometry error it claims to tolerate, since the paper currently probes only rotation perturbations.","supporting_citations":[{"cited_title":"LeanGaussian: Breaking pixel or point cloud correspondence in modeling 3D gaussians","cited_arxiv_id":null,"evidence_quote":"Supplies the query-based Gaussian decoder with deformable attention that FlexSplat extends from the single-view posed setting to uncalibrated multi-view."},{"cited_title":"UniGS: Modeling unitary 3D gaussians for novel view synthesis from sparse-view images","cited_arxiv_id":null,"evidence_quote":"Defines the posed multi-view unitary-Gaussian baseline with multi-view deformable cross-attention; FlexSplat inherits its decoder and compares directly against it on GSO."},{"cited_title":"VGGT: Visual geometry grounded transformer","cited_arxiv_id":null,"evidence_quote":"The geometry transformer used as the front-end, predicting per-image cameras and depth and jointly fine-tuned in FlexSplat."},{"cited_title":"3D gaussian splatting for real-time radiance field rendering.ACM Trans","cited_arxiv_id":null,"evidence_quote":"The 3D Gaussian splatting representation and differentiable renderer underlying the whole pipeline."},{"cited_title":"Splatter image: Ultra-fast single-view 3D reconstruction","cited_arxiv_id":null,"evidence_quote":"Source of the depth-guided mean parameterization (Eq. 2) and a pixel-aligned baseline whose primitive count grows with resolution."},{"cited_title":"No pose, no problem: Surprisingly simple 3D gaussian splats from sparse unposed images","cited_arxiv_id":null,"evidence_quote":"Provides the standard global-similarity alignment used to evaluate pose-free novel-view synthesis, allowing comparison against posed methods."}],"review_version":1}