{"id":"e2bc092f-c9b5-44f5-8e67-019e62139eae","arxiv_id":"2505.15185","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A feed-forward architecture that reuses a frozen depth foundation model to predict 3D Gaussian primitives, improving novel view synthesis and cross-dataset generalization.","lead":"MonoSplat turns a frozen monocular depth estimation model into a generalizable 3D Gaussian splatting system that reconstructs a scene from a few posed images and renders new views without per-scene training. The method reports state-of-the-art rendering quality on two benchmarks and better zero-shot generalization to unseen scene types than prior feed-forward 3D Gaussian splatting models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cross-dataset zero-shot evidence in Table 2 is confounded by target-specific depth bounds: Supplementary A.3 gives DTU bounds (2.125–4.525) while Sec. 4.1 states the default (0.5–100), and the paper does not state which range was used for MonoSplat versus baselines in Re10k→DTU.","rationale":"I read the paper in good faith: the architecture is coherent, the in-domain results are internally consistent with the reported training schedule, and Table 4 does provide partial support for the claim that depth-specific frozen features help cross-domain transfer. The load-bearing part of the central claim is generalization, and its strongest quantitative evidence is the DTU row of Table 2. That row is vulnerable because Supplementary A.3 supplies a DTU-specific depth interval that differs from the default interval in Sec. 4.1, while the main paper does not state which interval was used for MonoSplat and for each baseline. This is an internal ambiguity, not merely a disagreement with external consensus, and it interacts directly with the plane-sweep mechanism in Eq. (4). The reader’s weakest-assumption analysis focuses on whether monocular features are the true cause of the gains; the depth-range protocol is a more elementary confound that should be checked first. A single controlled rerun with unified depth grids would settle whether the DTU generalization claim survives. If it does, the reader’s CONDITIONAL verdict stands as-is; if it does not, the generalization claim would need substantial reinterpretation. I therefore keep the reader’s verdict unchanged rather than moving to ACCEPT or REJECT on the current evidence.","tokens_in":13956,"tokens_out":9004,"duration_ms":82794,"concrete_test":"Rerun the Re10k→DTU comparison with one shared depth grid for all methods under two conditions: (a) [0.5, 100] with D=128, and (b) [2.125, 4.525] with D=128. If MonoSplat’s DTU PSNR/SSIM advantage over MVSplat mostly disappears under condition (a), then the claimed zero-shot generalization gain is substantially attributable to target-specific depth-range knowledge rather than to the monocular depth features. If the advantage persists under both protocols, the concern is resolved. Report the same controlled comparison on Re10k→ACID, where no separately optimized narrow range exists.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The Abstract claims 'superior reconstruction quality and generalization capability,' and the decisive evidence for the generalization half is Table 2, especially Re10k→DTU (15.25 PSNR / 0.605 SSIM vs. MVSplat’s 13.94 / 0.473). The paper never specifies the depth range used in this zero-shot evaluation. Main-text Sec. 4.1 fixes the cost-volume range as d_near=0.5, d_far=100 with D=128 planes, but Supplementary A.3 states 'For the DTU dataset, we utilize the depth bounds of 2.125 and 4.525.' Since MonoSplat is never trained on DTU, this line can only refer to evaluation. The plane-sweep cost volume in Eq. (4) is strongly sensitive to the sampled depth interval: narrowing from 0.5–100 to 2.125–4.525 shrinks the search span from 99.5 to 2.4 units and injects knowledge about the target dataset’s depth distribution. If MonoSplat used the narrow DTU range while pixelSplat/MVSplat used the default range, part or most of the claimed 1.31 dB DTU gain could be an evaluation-protocol artifact rather than evidence for frozen monocular depth priors. The reader’s alternative concern (the confounded 'w/o MF' ablation) is real but partly mitigated by Table 4’s DINOv2/UniMatch comparisons; the depth-range protocol is unaddressed and should be settled first.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces MonoSplat, a feed-forward generalizable 3D Gaussian splatting model that builds on a frozen monocular depth foundation model (Depth Anything V2). A Mono-Multi Feature Adapter (DPT fusion plus cross-view Swin attention) converts monocular features into multi-view features, and an Integrated Gaussian Prediction module fuses these with a plane-sweep cost volume to predict Gaussian primitives. The model is trained with L_mse + lambda_lpips * L_lpips on RealEstate10K and ACID and evaluated zero-shot on DTU and ACID. The authors report state-of-the-art PSNR/SSIM/LPIPS on both main benchmarks, improved cross-dataset generalization, and efficiency with 10.3M trainable parameters, with code released.","tokens_in":14294,"tokens_out":4750,"duration_ms":37479,"significance":"If the results are reproducible, the paper demonstrates a practical and conceptually clean way to leverage frozen depth foundation models for generalizable Gaussian splatting, with strong cross-domain transfer and a low trainable parameter count. The ablation study comparing DINOv2, UniMatch, and Depth Anything V2 backbones is informative and supports the claim that depth-specific features matter. The main concerns are the ambiguous depth-range protocol in the zero-shot DTU evaluation and several internal inconsistencies, which must be resolved before the state-of-the-art generalization claim can be accepted.","major_comments":[{"comment":"Section 4.1 fixes the cost-volume depth range as d_near=0.5 and d_far=100 with D=128 planes, but Supplementary A.3 states 'For the DTU dataset, we utilize the depth bounds of 2.125 and 4.525.' Since MonoSplat is never trained on DTU, this sentence can only refer to the evaluation protocol, and the paper does not state whether the baselines in Table 2 (pixelSplat, MVSplat) used the same narrow DTU-specific range or the default 0.5-100 range. Because the plane-sweep cost volume in Eq. (4) is highly sensitive to the sampled depth interval, narrowing it by a factor of about 40 injects knowledge of the target dataset's depth distribution. The claimed 1.31 dB PSNR gain on Re10k-to-DTU (15.25 vs. 13.94) could therefore be an evaluation-protocol artifact rather than evidence for the frozen monocular depth prior. Please specify the exact depth range used for every method in Table 2 and, if the ranges differ, rerun the baselines under identical conditions.","section":"Sec. 4.1 and Supp. A.3, Table 2"},{"comment":"Section 4.3 states 'all models are trained for 20,000 iterations with a batch size of 14,' while Table 3's caption says 'models trained solely on Re10K with 200k iterations.' These two statements are mutually inconsistent, and the discrepancy directly affects the comparability of the ablations with the main results (300k iterations in Sec. 4.1). Please correct the number and report the actual training budget used for the ablations.","section":"Sec. 4.3 and Table 3 caption"},{"comment":"The Related Work discusses DepthSplat [48] as a concurrent method that also leverages Depth Anything features for Gaussian splatting, noting that its dual-branch design is parameter-inefficient. However, DepthSplat is absent from all quantitative comparisons in Tables 1 and 2 and from the efficiency analysis. Given the paper's claim of state-of-the-art quality and efficiency, the omission of a direct comparison (or at least a quantitative quote from the DepthSplat paper) leaves the central claim incomplete.","section":"Related Work and Tables 1-2"},{"comment":"Table 1 shows only small margins over MVSplat (e.g., 26.68 vs. 26.39 PSNR on RealEstate10K; 28.63 vs. 28.25 on ACID), yet no error bars, confidence intervals, or repeated-run statistics are provided. Given the 'convincingly demonstrate' language in the Abstract, please report variance across at least three runs or otherwise justify that the observed differences are statistically significant.","section":"Table 1 and Abstract"}],"minor_comments":[{"comment":"The caption contains a typo: 'Quantative' should be 'Quantitative', and 'boldedand underlined' is missing spaces.","section":"Table 2 caption"},{"comment":"The caption says 'tested on DTU [55]', but reference [55] is the RealEstate10K paper; DTU is [12]. Please correct the citation.","section":"Figure 6 caption"},{"comment":"References [12] and [50] appear to have formatting artifacts (a trailing '1' in the page field); please check the bibliography entries.","section":"References"},{"comment":"The sentence about DTU depth bounds should explicitly state whether these bounds are used during training, evaluation, or both, and for which methods, since the main text only defines the 0.5-100 range.","section":"Supplementary A.3"},{"comment":"The phrase 'the decoder's last layer' is ambiguous because Depth Anything V2 has multiple decoder outputs; please specify which output is used (e.g., the feature before the final depth head or the depth head's output).","section":"Sec. 3.1.2"}],"recommendation":"major_revision","confidential_remarks":"The depth-range ambiguity in the Re10k-to-DTU evaluation is the main blocker and should be the first point for the authors to address. If the baselines were evaluated with the default 0.5-100 range while MonoSplat used the DTU-specific 2.125-4.525 range, the main generalization claim collapses to a protocol artifact. The missing DepthSplat comparison is also significant given how closely that concurrent work resembles this one; the authors should at least report its published numbers if code is not available."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely useful paper in a crowded subfield, and the experiment that carries the \"generalization\" claim has an evaluation-protocol ambiguity that should be resolved before you trust the headline numbers.\n\nWhat's new: MonoSplat takes a frozen Depth Anything encoder, runs its multi-scale features through a DPT decoder and a lightweight cross-view transformer, and feeds both the resulting multi-view features and the monocular decoder features into the cost volume and the Gaussian refinement heads. That is a clean reuse of a depth foundation model rather than a learned cost-volume or epipolar network. It reports SOTA PSNR/SSIM/LPIPS on RealEstate10K and ACID with 10.3M trainable parameters, and the in-domain gains over MVSplat are consistent across both datasets. The ablation showing that fine-tuning the backbone hurts generalization by 2.6 dB is a useful counterpoint to the usual full-fine-tuning reflex, and Table 4's comparison against DINOv2/UniMatch backbones gives some evidence that the depth features specifically are doing the work.\n\nSoft spots. The one I'd push on first: the cross-dataset claim depends on Table 2, and the DTU row is potentially confounded by depth bounds. Main text fixes the plane-sweep range at 0.5-100, but Supplementary A.3 says \"for the DTU dataset, we utilize the depth bounds of 2.125 and 4.525\" — and since MonoSplat is never trained on DTU, that line can only refer to evaluation. If MonoSplat's cost volume samples the narrow range while pixelSplat and MVSplat sample 0.5-100, part of the 1.31 dB gain could be a protocol artifact. This needs a clear statement of which depth range each method used.\n\nSecond, the \"w/o MF\" ablation removes the monocular branch and the fusion subnetwork together, so the drop conflates removing the feature with cutting capacity. Table 4 partly mitigates this, but an ablation that keeps the fusion network and swaps in a frozen, non-depth backbone would be cleaner.\n\nMinor: Sec. 4.3 says 20k iterations while Table 3's caption says 200k; no error bars; and no experimental comparison with concurrent DepthSplat, which uses the same Depth Anything features in a dual-branch design. The related-work paragraph dismisses DepthSplat as parameter-redundant, but no number backs that.\n\nBottom line: the architecture is real, the in-domain results are solid, and the generalization story is plausible but not yet clean. Whoever referees it should ask for the depth-range protocol and the ablation detail; the code is released, so it is checkable. I'd send it out.","headline":"A clean and credible integration of a frozen monocular depth model into generalizable 3DGS, but the zero-shot DTU numbers hinge on a depth-range protocol ambiguity the authors must clear up.","tokens_in":14837,"tokens_out":4249,"would_cite":true,"duration_ms":33051,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A frozen monocular depth model, lightly adapted, yields state-of-the-art generalizable 3D Gaussian splatting, the paper claims.","keywords":["3D Gaussian Splatting","generalizable novel view synthesis","monocular depth foundation model","zero-shot cross-dataset generalization","plane-sweep cost volume","feature adapter","Depth Anything","feed-forward 3D reconstruction"],"falsifier":"A decisive test would be to replace the frozen depth encoder's features with randomly initialized features of identical shape, keeping every other component fixed, and then measure zero-shot PSNR on DTU after training on RealEstate10K. If the score does not drop substantially below the reported 15.25 dB, the monocular priors are doing little of the work; if it collapses toward MVSplat's 13.94 dB, the priors are the cause. A second observation: run the same protocol on strongly out-of-distribution imagery such as medical or satellite data, where the depth foundation model has no pretraining; if the advantage disappears there, the generalization claim is bounded by the pretraining distribution.","tokens_in":13753,"feed_emoji":"🎥","tokens_out":9209,"duration_ms":70947,"temperature":0.7,"pith_summary":"The paper claims that a generalizable 3D Gaussian splatting model can be built by freezing a monocular depth foundation model and adapting its features, instead of training a reconstruction network from scratch. The proposed system, MonoSplat, converts Depth Anything V2 features into multi-view representations with a lightweight attention adapter, then fuses them into the cost volume and refinement network that predicts Gaussian primitives. On RealEstate10K and ACID it reports the best PSNR, SSIM, and LPIPS among compared feed-forward methods, and in zero-shot transfer from RealEstate10K to DTU and ACID it reports larger gains than pixelSplat and MVSplat. If the claim holds, rich monocular depth priors are a reusable, parameter-efficient ingredient for real-time 3D reconstruction of novel scenes without per-scene optimization.","feed_headline":"Frozen depth priors lift 3D Gaussian splatting past prior models","feed_subtitle":"Adapts frozen Depth Anything features into 3D Gaussians with 10.3M trainable parameters for zero-shot view synthesis.","key_machinery":"The load-bearing mechanism is the Mono-Multi Feature Adapter combined with the Integrated Gaussian Prediction module. The Adapter first passes multi-scale encoder features from the frozen Depth Anything V2 backbone through a DPT head to form a unified per-view feature, then runs local-window cross-view attention in the style of Swin Transformer so that each view's features incorporate information from its nearest neighboring views. The Integrated Gaussian Prediction module feeds both these multi-view features and the frozen decoder's monocular features into a plane-sweep cost volume, refines the resulting depth distribution, and decodes Gaussian parameters through depth and appearance heads. The frozen backbone is what preserves the pretrained geometric priors and keeps trainable parameters low, and the paper argues this is why the model transfers across domains.","core_discovery":"MonoSplat's central claim is that a frozen monocular depth foundation model can be repurposed as the backbone of a generalizable 3D Gaussian splatting system. The paper introduces a Mono-Multi Feature Adapter that fuses multi-scale encoder features through a DPT head and then applies Swin-based cross-view attention, turning view-specific monocular features into geometry-aware multi-view features. These are combined with the decoder's monocular features in an integrated cost volume and in the refinement network that predicts Gaussian positions, opacities, covariances, and colors. On RealEstate10K and ACID, the method reports the best PSNR, SSIM, and LPIPS among the compared feed-forward approaches, and in zero-shot transfer from RealEstate10K to DTU and ACID it reports larger gains than pixelSplat and MVSplat while using only 10.3M trainable parameters out of 30.3M total.","pith_inferences":["If the monocular depth priors are the active ingredient, the same recipe could transfer to other feed-forward reconstruction backbones: swapping in a different pretrained depth model should shift the generalization boundary accordingly.","The paper's ablation removes the monocular branch together with the fusion subnetwork, so part of the reported drop could reflect reduced capacity; a test that keeps capacity constant would isolate the priors themselves.","The frozen-encoder design points to a practical pattern for other 3D tasks: adapt a large pretrained 2D backbone with a small cross-view layer rather than fine-tuning it, preserving broad priors at low parameter cost.","The cross-dataset gains are demonstrated only on natural scenes; on imagery far outside the depth model's pretraining distribution, such as medical or satellite data, the advantage may shrink or vanish."],"forward_implications":["MonoSplat can render novel views of unseen scenes from two posed images without per-scene optimization, with inference times close to those of MVSplat.","Zero-shot transfer from indoor RealEstate10K to DTU and ACID improves over prior feed-forward splatting models, and the improvement grows as the domain gap widens.","Keeping the depth backbone frozen preserves geometric priors; fine-tuning the backbone drops cross-domain PSNR by 2.61 dB in the paper's ablation.","The full model uses 30.3M parameters with only 10.3M trainable, and runs in 0.857 GB of GPU memory, comparable to or better than the cost-volume baseline.","Larger frozen Depth Anything V2 variants monotonically improve both in-domain quality and cross-domain generalization."],"supporting_citations":[{"why":"Supplies the frozen monocular depth foundation model (Depth Anything V2) whose features carry the geometric priors.","marker":"[50]"},{"why":"Provides the DPT decoder design used to fuse the encoder's multi-scale features into a single representation.","marker":"[29]"},{"why":"Baseline cost-volume-based Gaussian splatting method that MonoSplat extends and compares against on both datasets.","marker":"[6]"},{"why":"Baseline epipolar-transformer method that pioneered feed-forward Gaussian splatting and provides the training protocol.","marker":"[1]"},{"why":"Supplies the local-window attention mechanism used for efficient cross-view feature aggregation.","marker":"[19]"},{"why":"Defines the 3D Gaussian splatting representation and differentiable rasterizer that the output primitives use.","marker":"[14]"},{"why":"Concurrent work connecting monocular depth features to Gaussian splatting, used to position MonoSplat's efficiency advantage.","marker":"[48]"},{"why":"RealEstate10K is the main training and in-domain evaluation dataset.","marker":"[55]"},{"why":"DTU is the primary out-of-distribution zero-shot generalization target.","marker":"[12]"},{"why":"ACID is the aerial benchmark for in-domain evaluation and one of the zero-shot transfer targets.","marker":"[15]"}],"fun_headline_variants":["MonoSplat: depth priors for zero-shot 3D Gaussian splatting","Frozen depth features power generalizable 3D Gaussian rendering","10.3M trainable params upend 3D view synthesis via depth priors","Adapter fuses depth priors into 3D Gaussians for robust novel views","MonoSplat generalizes 3D splatting with monocular depth foundation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central load-bearing assumption is that monocular depth features learned on large-scale natural images transfer to arbitrary unseen scenes and genuinely improve cost-volume matching, so that the measured cross-dataset gains come from those priors rather than from extra model capacity or dataset-specific artifacts.","fun_headline_variants_meta":{"raw":{"variants":["MonoSplat: depth priors for zero-shot 3D Gaussian splatting","Frozen depth features power generalizable 3D Gaussian rendering","10.3M trainable params upend 3D view synthesis via depth priors","Adapter fuses depth priors into 3D Gaussians for robust novel views","MonoSplat generalizes 3D splatting with monocular depth foundation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000375,"raw_usage":{"total_tokens":2002,"prompt_tokens":951,"completion_tokens":1051,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":944}},"tokens_in":567,"tokens_out":1051,"duration_ms":8126,"temperature":1.0,"reasoning_tokens":944,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:21:54.829826+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test would be to replace the frozen depth encoder's features with randomly initialized features of identical shape, keeping every other component fixed, and then measure zero-shot PSNR on DTU after training on RealEstate10K. If the score does not drop substantially below the reported 15.25 dB, the monocular priors are doing little of the work; if it collapses toward MVSplat's 13.94 dB, the priors are the cause. A second observation: run the same protocol on strongly out-of-distribution imagery such as medical or satellite data, where the depth foundation model has no pretraining; if the advantage disappears there, the generalization claim is bounded by the pretraining distribution.","supporting_citations":[{"cited_title":"Depth anything: Unleashing the power of large-scale unlabeled data","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen monocular depth foundation model (Depth Anything V2) whose features carry the geometric priors."},{"cited_title":"pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction","cited_arxiv_id":null,"evidence_quote":"Baseline epipolar-transformer method that pioneered feed-forward Gaussian splatting and provides the training protocol."},{"cited_title":"Swin transformer: Hierarchical vision transformer using shifted windows","cited_arxiv_id":null,"evidence_quote":"Supplies the local-window attention mechanism used for efficient cross-view feature aggregation."},{"cited_title":"3d gaussian splatting for real-time radiance field rendering.TOG, 42(4), 2023","cited_arxiv_id":null,"evidence_quote":"Defines the 3D Gaussian splatting representation and differentiable rasterizer that the output primitives use."},{"cited_title":"Stereo magnification: learning view syn- thesis using multiplane images.TOG, page 65, 2018","cited_arxiv_id":null,"evidence_quote":"RealEstate10K is the main training and in-domain evaluation dataset."},{"cited_title":"Large scale multi-view stereopsis evalu- ation","cited_arxiv_id":null,"evidence_quote":"DTU is the primary out-of-distribution zero-shot generalization target."},{"cited_title":"Infinite nature: Perpetual view generation of natural scenes from a single image","cited_arxiv_id":null,"evidence_quote":"ACID is the aerial benchmark for in-domain evaluation and one of the zero-shot transfer targets."}],"review_version":1}