{"id":"6ca2614b-7862-4461-83ec-a395d715a032","arxiv_id":"2509.17246","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"SPFSplatV2 reconstructs 3D Gaussian scenes and camera poses from sparse unposed images with a single shared transformer backbone, using masked attention and a reprojection loss, and reports state-of-the-art results without ground-truth pose supervision.","lead":"This paper introduces a 3D scene reconstruction method that works from a few photos without needing camera positions during training or use, predicting both the 3D scene and camera poses from images alone. It reports state-of-the-art novel view synthesis and pose accuracy, which could let future systems train on vast collections of unposed photos and videos.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'self-supervised' claim is narrower than stated: most of the reported performance is inherited from pose/geometry-supervised MASt3R or VGGT initialization, and random initialization requires a DUSt3R distillation warm-up, so the method does not learn geometry from image-level losses alone.","rationale":"The reader's weakest_assumption identifies exactly the concern I consider most load-bearing: the method's performance is largely inherited from pose/geometry-supervised pretrained models (MASt3R/VGGT), and random initialization requires a DUSt3R distillation warm-up that is a form of geometric supervision. This does not make the reported numbers false, but it substantially narrows the abstract's claim that the method requires 'no ground-truth poses' and is 'self-supervised.' The paper is transparent about this in its Limitations section and Table XI, which is why the concern is not disqualifying. The conditionality is appropriate: the claims should be rephrased or the evaluation should be augmented to separate the contribution of the pretrained backbone from the pose-free training objective. My recommended concrete test—training from random initialization without any distillation—would settle whether the training framework alone can learn 3D geometry from image-level losses, which is the crux of the central claim. Until then, the verdict should remain CONDITIONAL, as the reader concluded. I see no reason to move to ACCEPT or REJECT: the evidence is internally consistent, the concern is acknowledged, and the proposed check is feasible.","tokens_in":24825,"tokens_out":17971,"duration_ms":155624,"concrete_test":"Train SPFSplatV2 from random initialization using only L_render (Eq. 13) and L_reproj (Eq. 14), omitting the DUSt3R distillation warm-up and any pretrained geometric weights, for the same training budget and schedule as in Section IV-B. Report whether training converges and the final RE10K NVS PSNR/SSIM/LPIPS. If training diverges, collapses, or stays below, say, 20 dB PSNR, it confirms that geometric supervision (pretraining or distillation) is necessary and that the 'self-supervised' label applies only to the fine-tuning stage, not to the training paradigm as a whole. Additionally, to isolate the effect of initialization, report the performance of the frozen MASt3R and VGGT backbones (no fine-tuning) under the same evaluation protocol, making explicit what fraction of the final SOTA numbers is inherited from the pretrained weights.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SPFSplatV2 is trained 'purely with image rendering and reprojection losses, without ground-truth poses' and still achieves SOTA. The paper's own Table XI shows this is only true for the fine-tuning stage: random initialization with a DUSt3R point-cloud distillation warm-up drops RE10K NVS* PSNR from 26.157 (MASt3R init) to 22.394, a 3.76 dB loss. The warm-up itself is a form of geometric supervision (distillation of a model trained with pose/depth supervision), and Section IV-D states it is 'essential' because 'training solely with a photometric loss, especially without ground-truth geometric supervision, makes it difficult for the network to learn Gaussians in the canonical space.' Section V (Limitations) explicitly acknowledges that the method 'benefits from the priors provided by supervised models such as MASt3R and VGGT.' Thus the abstract's 'absence of pose supervision' is technically about the loss objective, but the model's geometric capability is largely inherited from pretrained, supervised reconstruction models. This narrows the substantive claim from 'self-supervised pose-free training suffices' to 'fine-tuning a supervised geometry backbone with a pose-free objective works well,' which is a materially weaker and different contribution. The concern is load-bearing because it directly affects how the central result should be interpreted and how much of the SOTA performance is attributable to the proposed training paradigm rather than to the initialization.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SPFSplatV2, a feed-forward framework for novel view synthesis from unposed sparse multi-view images using 3D Gaussian splatting. The method uses a shared ViT encoder and a masked multi-view decoder to jointly predict pixel-aligned Gaussians in a canonical space and relative camera poses, trained with a rendering loss and a reprojection loss, without any ground-truth pose supervision in the objective. Two variants are presented: SPFSplatV2 built on a MASt3R-style architecture and SPFSplatV2-L built on a VGGT-style architecture. Experiments on RealEstate10K, ACID, DTU, DL3DV, and ScanNet++ report state-of-the-art results in novel view synthesis, cross-dataset generalization, and relative pose estimation, together with efficiency and ablation studies, including an ablation that compares against training with ground-truth poses and an initialization analysis.","tokens_in":25028,"tokens_out":7965,"duration_ms":62273,"significance":"The reported results, if they hold, would be a meaningful advance for pose-free generalizable 3D reconstruction: the masked-attention single-branch design is a clean solution to the target-leakage problem, the reprojection loss is shown to be essential for stability, and the framework's compatibility with two different reconstruction backbones suggests the training paradigm is portable. The paper is also commendably transparent in running the ground-truth-pose and initialization ablations in Tables X and XI, which most pose-free papers would omit. The main significance caveat is that the method delivers its strongest performance when initialized from MASt3R or VGGT, models that were themselves trained with pose or depth supervision; the contribution is therefore best understood as a pose-free fine-tuning objective for strong geometric backbones, not as evidence that such geometry can be learned from image-level losses alone.","major_comments":[{"comment":"The abstract and Section IV-C state that the method works 'despite the absence of pose supervision' and 'despite no geometry priors during training,' but Section V concedes that the method 'benefits from the priors provided by supervised models such as MASt3R and VGGT,' and Table XI shows that random initialization with a DUSt3R distillation warm-up drops RE10K PSNR from 26.157 to 22.394. In addition, Section IV-D states that training solely with a photometric loss without ground-truth geometric supervision makes it difficult to learn Gaussians in canonical space, and that the warm-up is 'essential.' These statements put the 'no pose supervision' claim in a materially narrower form: the method fine-tunes pose- or depth-supervised reconstruction backbones with a pose-free objective. I recommend that the claims in the abstract, Section IV-C, and the conclusion be reworded to state exactly this, and that the contribution be framed as a pose-free fine-tuning paradigm rather than purely image-supervised training from scratch.","section":"§IV-C, §IV-D, §V, Table XI"},{"comment":"The random-initialization ablation is not a clean test of fully self-supervised training because the reported 'Random' row uses a warm-up phase with a DUSt3R point-cloud distillation loss for the first 10,000 steps. As the text itself says, this supervision is essential; without it the photometric-only objective fails to learn canonical Gaussians. The table and surrounding text should therefore label this setting as 'random init + DUSt3R distillation warm-up,' and the limitation should be acknowledged in the conclusion. As reported, the row conflates two effects: removal of pretrained weights and addition of geometric distillation, so it does not by itself establish what a purely image-supervised run would achieve.","section":"§IV-D, Table XI"},{"comment":"All quantitative results are reported as single numbers with no standard errors, confidence intervals, or repeated-seed information. Several headline comparisons are very close: for example, Table I lists SPFSplatV2* at 26.157 dB versus SPFSplat* at 25.845 dB, and Table II lists SPFSplatV2* at 26.809 dB versus SPFSplat* at 26.796 dB. Given that the central claim is state-of-the-art performance, the paper should at least report results over multiple seeds with variance, or otherwise justify that the differences are not within run-to-run noise. This is particularly important for the zero-shot cross-dataset tables, which are reported on what appear to be single evaluation passes.","section":"§IV-C, Tables I–IV"}],"minor_comments":[{"comment":"In the Splatt3R row, the average SSIM and LPIPS values (0.337 and 0.596) appear to be swapped relative to the overlap-specific columns; please check and correct.","section":"Table I"},{"comment":"The sentence 'Despite no geometry priors during training' is inconsistent with the initialization discussion in Section IV-D and Section V; please revise it to avoid contradicting the paper's own limitation statement.","section":"§IV-C, Relative Pose Estimation"},{"comment":"The statement 'SPFSplatV2-L consistently outperforms SPFSplatV2 both with and without pose alignment' is not supported by Table III on ACID with pose alignment, where SPFSplatV2* reports PSNR 26.802 versus SPFSplatV2-L* 26.680; please qualify the claim by metric or dataset.","section":"§IV-C, Cross-Dataset Generalization"},{"comment":"The phrase 'according to the architecuture of CroCo' contains a typo: 'architecuture' should be 'architecture.'","section":"§II, Related Work"},{"comment":"Please state explicitly whether the multi-view dropout strategy is applied in all main experiments and ablations, since it affects the comparability of the reported numbers.","section":"§IV-B, Implementation Details"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid engineering contribution that likely belongs in this venue after revision. The main issue is terminology: the authors use 'self-supervised' to mean 'no ground-truth pose loss,' which is defensible, but they occasionally overclaim by saying 'no geometry priors.' Please encourage the authors to distinguish 'no ground-truth pose labels in the loss' from 'no geometric supervision anywhere,' since the latter is false given the MASt3R/VGGT initialization and the DUSt3R warm-up. I do not see grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a competent, incremental-but-real improvement to the authors' own SPFSplat. The masked attention decoder, learnable pose tokens and multi-view dropout each earn their keep, and the VGGT variant shows the training paradigm transfers across backbones. Empirically it is the best pose-free sparse-view NVS method on the benchmarks tested, including beating NoPoSplat which trains with ground-truth poses. That result is worth taking seriously.\n\nThe main thing I would push back on is the framing. The abstract says \"despite the absence of pose supervision\" and calls the method self-supervised. That is only true for the fine-tuning stage. Both variants initialize from models trained with explicit geometric supervision (MASt3R, VGGT), and Table XI shows random initialization plus a DUSt3R point-cloud distillation warm-up is \"essential\" and costs 3.76 dB. The Limitations section admits the method \"benefits from the priors provided by supervised models.\" The honest claim is: given a supervised geometry backbone, you can train a pose-free head and rendering objective that matches pose-supervised methods. That is a meaningful and practical result, but it is not \"self-supervised from scratch.\" The paper should say that in the abstract, not bury it in the ablations.\n\nOther soft spots are minor. No error bars or significance tests, and some SOTA margins are small. Code and models are promised but not available. Table I has a swapped SSIM/LPIPS average in the Splatt3R row. None of these undercut the main result.\n\nThe strongest evidence is Table X: adding a ground-truth pose loss during training improves pose AUC but barely changes NVS, which supports the claim that the pose-free objective is not just riding on the initialization. The cross-dataset numbers are broad enough to show generalization, not just RE10K overfitting.\n\nWho is this for? Anyone working on generalizable 3D reconstruction or pose-free NVS. It deserves a serious referee. I recommend accepting after the authors revise the \"self-supervised\" claim, release code, and add multi-seed results or error bars for the main tables.","headline":"Solid incremental extension of SPFSplat with strong empirical results, but the 'self-supervised' claim outruns the evidence: most performance comes from supervised MASt3R/VGGT initialization.","tokens_in":25647,"tokens_out":2401,"would_cite":true,"duration_ms":21236,"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":"SPFSplatV2 claims that a feed-forward 3D Gaussian splatting model trained without ground-truth camera poses—using only rendering and reprojection losses—matches or beats pose-required and supervised pose-free methods on novel view…","keywords":["3D Gaussian splatting","self-supervised learning","pose-free novel view synthesis","sparse-view reconstruction","camera pose estimation","masked attention","reprojection loss","feed-forward reconstruction"],"falsifier":"Train the full pipeline from randomly initialized weights with the DUSt3R point-cloud distillation warm-up deleted, leaving only the rendering and reprojection losses; if the model then collapses to near-identity poses and blurred renderings (pose AUC near zero), the claim that ground-truth poses are unnecessary for this training objective is falsified for the from-scratch regime.","tokens_in":24503,"feed_emoji":"🎥","tokens_out":5739,"duration_ms":44002,"temperature":0.7,"pith_summary":"The paper tries to establish that ground-truth camera poses are not needed to train a feed-forward 3D reconstruction model for novel view synthesis. It proposes a single network that, given a few unposed images, predicts both 3D Gaussian primitives and camera poses in one shared canonical space, and is trained by rendering target views with the estimated poses plus a pixel-reprojection loss. The authors claim this self-supervised, pose-free pipeline is stable enough to beat pose-required methods such as pixelSplat and MVSplat, supervised pose-free methods such as NoPoSplat, and earlier self-supervised pose-free methods, both in-domain and out-of-domain, while also producing competitive relative camera poses. If true, the result matters because it removes the main bottleneck to training on large, diverse, unannotated video and photo collections.","feed_headline":"Pose-free 3D model outdoes pose-supervised view synthesis","feed_subtitle":"Trained only on image and reprojection losses, it still wins on novel view benchmarks.","key_machinery":"The load-bearing mechanism is a masked multi-view decoder with learnable pose tokens. During training, target images enter the same forward pass as context images, but cross-attention masks prevent context tokens from seeing target tokens, so the reconstructed Gaussians depend only on the context views; target tokens attend to all views so the pose head can estimate the target pose. The second mechanism is the reprojection loss, which projects each predicted Gaussian center back into its own image with the estimated pose and penalizes pixel displacement, acting as a differentiable geometric constraint without ground-truth poses.","core_discovery":"The central discovery is that joint optimization of 3D Gaussians and camera poses, with a shared backbone and a masked attention decoder, makes self-supervised pose-free training practical for sparse views. By letting context tokens attend only to context tokens while target tokens attend to everything, the model prevents target-view information from leaking into the reconstruction, yet still uses the target image to estimate its pose for the rendering loss. A pixel-wise reprojection loss on the predicted Gaussian centers and context poses supplies the geometric constraint that keeps the Gaussians pixel-aligned and the training stable. The paper reports that the resulting SPFSplatV2 and the larger VGGT-based SPFSplatV2-L outperform prior work across overlap regimes on RealEstate10K and ACID, generalize zero-shot to unseen datasets, and surpass many geometric-supervision methods in relative pose estimation.","pith_inferences":["Editorial inference: the strongest reading of the results is not that poses are unnecessary, but that the rendering-plus-reprojection objective can replace explicit pose labels during fine-tuning, provided the network starts from weights that already encode geometric priors.","Editorial inference: the masked-attention design may be the key reusable idea beyond splatting—any feed-forward reconstruction task that needs target-view information for supervision but not for the representation could use the same attention mask.","Editorial inference: a direct testable extension is to train the same pipeline on an unposed dataset that has no SfM poses at all, using only the photometric and reprojection losses, and measure whether the pose-estimation gains in Table VIII persist or saturate."],"forward_implications":["Training no longer requires SfM pose preprocessing, so the same framework can be trained directly on larger and more diverse unposed video and photo collections; the paper shows that adding DL3DV to training improves pose estimation on multiple benchmarks.","The training paradigm transfers across reconstruction architectures: the same losses and masked-attention scheme work on a MASt3R-style backbone and a VGGT-style backbone, suggesting the self-supervised objective, not a specific network design, is what removes the pose dependency.","Because estimated poses and reconstructed Gaussians are jointly optimized, rendering with the predicted poses is competitive with methods that use evaluation-time pose alignment, which decouples rendering quality from pose accuracy.","A single model with multi-view dropout handles different numbers and spatial arrangements of context views, so practitioners do not need to train separate models for two-view versus many-view inputs."],"supporting_citations":[{"why":"The prior SPFSplat method that this paper extends; supplies the self-supervised pose-free baseline and the two-branch design that masked attention replaces.","marker":"[22]"},{"why":"MASt3R provides the pretrained encoder, decoder, and Gaussian center head weights that initialize SPFSplatV2, and the asymmetric decoder structure it adapts.","marker":"[23]"},{"why":"VGGT provides the pretrained architecture and weights for SPFSplatV2-L, including the unified decoder and two learnable pose tokens.","marker":"[24]"},{"why":"NoPoSplat is the main supervised pose-free baseline, the source of the evaluation protocol (overlap split, EPA) and the intrinsics-embedding idea.","marker":"[7]"},{"why":"pixelSplat is a pose-required baseline that the paper must beat, and a source of the epipolar-transformer and pixel-aligned Gaussian design ideas.","marker":"[5]"},{"why":"MVSplat is the other pose-required baseline and a comparison point for efficiency and novel view synthesis quality.","marker":"[6]"},{"why":"SelfSplat is the closest self-supervised pose-free competitor using cross-view U-Nets; the paper argues separate pose and reconstruction modules cause its weaknesses.","marker":"[21]"},{"why":"PF3plat is the other self-supervised pose-free baseline, representing the inefficient off-the-shelf RANSAC-based pipeline the paper contrasts with.","marker":"[20]"},{"why":"DUSt3R provides the random-initialization warm-up distillation loss in the initialization ablation and is a geometric SfM baseline for pose estimation.","marker":"[29]"}],"fun_headline_variants":["Self-supervised 3D Gaussians outdo pose-supervised rivals","No poses needed: SPFSplatV2 hits SOTA on sparse views","Pose-free splatting beats supervised methods in view synthesis","SPFSplatV2: Self-supervised splatting without pose labels","Learn 3D scene and poses from just sparse unposed images"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that training is pose-free rests on the assumption that initializing from MASt3R or VGGT pretrained weights—models trained with pose or depth supervision—does not count as using ground-truth poses; Table XI shows most of the performance disappears under random initialization.","fun_headline_variants_meta":{"raw":{"variants":["Self-supervised 3D Gaussians outdo pose-supervised rivals","No poses needed: SPFSplatV2 hits SOTA on sparse views","Pose-free splatting beats supervised methods in view synthesis","SPFSplatV2: Self-supervised splatting without pose labels","Learn 3D scene and poses from just sparse unposed images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000235,"raw_usage":{"total_tokens":1490,"prompt_tokens":923,"completion_tokens":567,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":472}},"tokens_in":539,"tokens_out":567,"duration_ms":5283,"temperature":1.0,"reasoning_tokens":472,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:47:57.123694+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the full pipeline from randomly initialized weights with the DUSt3R point-cloud distillation warm-up deleted, leaving only the rendering and reprojection losses; if the model then collapses to near-identity poses and blurred renderings (pose AUC near zero), the claim that ground-truth poses are unnecessary for this training objective is falsified for the from-scratch regime.","supporting_citations":[{"cited_title":"No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images,","cited_arxiv_id":null,"evidence_quote":"NoPoSplat is the main supervised pose-free baseline, the source of the evaluation protocol (overlap split, EPA) and the intrinsics-embedding idea."},{"cited_title":"Selfsplat: Pose-free and 3d prior-free generalizable 3d gaussian splat- ting,","cited_arxiv_id":null,"evidence_quote":"SelfSplat is the closest self-supervised pose-free competitor using cross-view U-Nets; the paper argues separate pose and reconstruction modules cause its weaknesses."}],"review_version":2}