Pith. sign in

REVIEW 3 major objections 5 minor 55 references

AirSplat claims two training-time fixes enable 3D vision foundation models to generate high-fidelity novel views from unposed images, achieving state-of-the-art results on RealEstate10K, DL3DV, and ACID.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

AirSplat fine-tunes a 3D vision foundation model with a self-corrected pose loop and teacher-rated opacity pruning to achieve state-of-the-art pose-free novel view synthesis.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection AirSplat's two training-time components are solid and the ablations are convincing, but the test-time pose alignment protocol is underspecified and must be clarified before the SOTA margins can be taken at face value. the 3 major comments →

arxiv 2603.25129 v2 pith:IAEV7UDM submitted 2026-03-26 cs.CV

AirSplat: Alignment and Rating for Robust Feed-Forward 3D Gaussian Splatting

Minh-Quan Viet Bui , Jaeho Moon , Munchurl Kim This is my paper
classification cs.CV
keywords pose-free novel view synthesis3D Gaussian splatting3D vision foundation modelsself-consistent pose alignmentrating-based opacity matchingmulti-view consistencypose-geometry discrepancyfeed-forward 3D reconstruction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

AirSplat argues that two specific training problems prevent 3D vision foundation models with view-synthesis heads from producing high-quality novel views from unposed images: a pose-geometry mismatch during training (predicted target poses are misaligned with the context-derived scene geometry) and local multi-view inconsistencies that create 'floater' artifacts. The proposed framework counters the first with Self-Consistent Pose Alignment (SCPA), a feedback loop that renders a proxy view, re-predicts the pose, measures the systematic drift between the two pose estimates, and applies its inverse to align supervision with the scene geometry. The second is addressed by Rating-based Opacity Matching (ROM), which uses a lightweight sparse-view teacher model to rate each Gaussian primitive's geometric consistency and then penalizes a primitive's opacity when it exceeds the teacher's rating, pruning inconsistent primitives. On RealEstate10K, DL3DV, and ACID, AirSplat reports large gains over prior pose-free methods and, on several settings, surpasses pose-required baselines. The central claim is that decoupling coordinate drift from photometric loss and adding teacher-driven local consistency feedback unlocks high-fidelity pose-free NVS while preserving the foundation model's geometry estimation.

Core claim

AirSplat's discovery is that the two obstacles—the pose-geometry discrepancy induced by the context-target training strategy and the multi-view inconsistency of predicted 3D Gaussian primitives—are the main reasons 3D vision foundation models with view-synthesis heads underperform in novel view synthesis. SCPA resolves the first by rendering a proxy image, re-predicting the target pose, computing the SE(3) drift between the two pose predictions, and applying its negation to anchor supervision to the context-derived geometry. ROM resolves the second by using a pre-trained sparse-view NVS teacher to rate each primitive's geometric error relative to the teacher's local consensus, then treats th

What carries the argument

The central mechanisms are Self-Consistent Pose Alignment (SCPA) and Rating-based Opacity Matching (ROM). SCPA is a training-time feedback loop: it renders the initial predicted target view, re-feeds that render along with the context views to re-predict the target pose, computes the relative SE(3) transformation between the initial and re-predicted poses via the Lie-algebra logarithm map, and applies the negated transformation to the initial pose before using it for photometric supervision. This effectively decouples the coordinate drift of the pose head from the geometry optimization. ROM uses a pre-trained, lightweight two-view feed-forward 3DGS model as a teacher oracle: for each predict

Load-bearing premise

The reported performance margins assume the same test-time pose optimization for ground-truth alignment is applied to AirSplat as to all pose-free baselines; the paper's Section 4.2 wording is ambiguous about whether AirSplat itself undergoes this alignment, so the comparison could be unfair if it is not applied identically.

What would settle it

Retrain or finetune a pose-free baseline with the identical evaluation protocol, applying test-time pose alignment to all methods including AirSplat and each baseline, and re-measure PSNR on RE10K at 12 views; if the margin over DA3 drops below the reported 2.3 dB, the central SOTA claim would be weakened. Alternatively, run AirSplat with and without test-time pose optimization and compare the gap to baselines run with and without it.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Pose-free feed-forward NVS can reach or exceed pose-required methods on dense benchmarks; e.g., on RE10K at 12 views, AirSplat reaches 23.08 PSNR versus 22.56 for DepthSplat, a pose-required baseline.
  • Training a unified 3D vision foundation model with both geometry and view-synthesis heads can improve NVS without degrading geometry estimation, since only the Gaussian head is fine-tuned.
  • The context-target training strategy, when combined with SCPA's drift correction, is a sound way to supervise novel viewpoint synthesis without external poses.
  • Teacher-driven geometric ratings can act as a general regularizer for feed-forward 3D reconstruction, potentially applicable to other primitive-based representations.
  • The method transfers zero-shot across datasets (e.g., ACID), suggesting the learned consistency and alignment priors are not overfitted to the training distribution.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the systematic pose drift observed by SCPA is a general property of context-target training across different 3DVFMs, a similar self-alignment module could improve other pose-free NVS baselines without retraining their core geometry heads.
  • The opacity-as-rating interpretation suggests that opacity in learned 3D representations can be read as a confidence or consistency signal; this could inform uncertainty-aware rendering in other feed-forward models.
  • Because the teacher in ROM is a lightweight two-view model, the quality of its ratings may bound the pruning quality; swapping in a stronger multi-view teacher could be a testable extension.
  • The paper's stated limitation that occluded or unobserved regions appear as voids suggests combining AirSplat with generative diffusion-based inpainting as a post-processing step could address its deterministic blind spots.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes AirSplat, a training framework for adapting 3D-VS-VFMs (specifically DA3-GIANT) to pose-free feed-forward 3D Gaussian splatting. The two main contributions are Self-Consistent Pose Alignment (SCPA), which corrects predicted target poses by re-predicting poses from initial renderings and applying the inverse SE(3) drift, and Rating-based Opacity Matching (ROM), which uses a pretrained sparse-view teacher model to assign geometric ratings that act as upper bounds on primitive opacity. The method is evaluated on RE10K, DL3DV, and ACID, reporting state-of-the-art pose-free NVS results and also outperforming several pose-required baselines. Ablations on RE10K and DL3DV support the individual contributions of SCPA and ROM.

Significance. If the evaluation protocol is symmetric and the teacher details are fully specified, this is a meaningful advance: it demonstrates that a 3D-VS-VFM can be fine-tuned to match or exceed dedicated pose-based feed-forward NVS systems on large-scale benchmarks while preserving the foundation model's geometry estimation (Table 6). The paper has several strengths: consistent ablations across two datasets, a zero-shot cross-dataset evaluation, a training-time overhead analysis, and a candid limitations section. However, the headline margins in Tables 1–3 rest directly on the test-time pose-alignment protocol, and the ROM module's use of a pose-requiring teacher is underspecified. Both points are load-bearing for the central claims.

major comments (3)
  1. [Sec. 4.2 (Evaluation Protocol)] The sentence 'we adopt the test-time pose optimization for ground truth alignment for all pose-free NVS baselines' is ambiguous as to whether AirSplat itself is also aligned. The state-of-the-art margins in Tables 1–3 are computed against pose-free baselines, so an asymmetric protocol — alignment applied only to baselines, or a different alignment for the proposed method — would invalidate the comparison. Please state explicitly that AirSplat is included in the same alignment procedure, and specify the optimizer objective (e.g., Sim(3) or SE(3) alignment of predicted poses to ground truth), the views used by the optimizer (context-only or also target frames), and the number of iterations. If the alignment uses ground-truth target poses or images, disclose this as a test-time adaptation step.
  2. [Sec. 3.3 (Rating-based Opacity Matching)] The teacher model \tilde{f}_\phi is cited as DepthSplat [46], which is not a pose-free model. The text never states what camera parameters are used when feeding the view pair (I_ctx,v, I_ctx,v') to the teacher, nor how the teacher's primitives are projected when computing \tilde{\epsilon}_j in Eq. (6). If ground-truth poses are used during training, that must be explicitly disclosed, because the method is advertised as pose-free; if predicted poses are used, the coordinate-frame relationship between the teacher's \tilde{G} and the student's G must be described. Without this information, Eqs. (6)–(7) are not reproducible and the claimed pose-free nature of the geometric ratings is incomplete. Also clarify how 'the exact same pixel' correspondence between student primitive g_i and teacher primitive \tilde{g}_j is established.
  3. [Sec. 4.2 / Tables 1–3 (baseline evaluation)] The manuscript does not state which implementation or official checkpoint was used for each baseline, nor whether all baseline numbers were produced with the same version of the test-time pose-alignment protocol. Since the SOTA claim depends on the fairness of the comparison, please include a reproducibility statement specifying the source of each baseline result and whether the evaluation code aligns all methods in exactly the same way. This is especially important for the ACID zero-shot results in Table 3, where the protocol description is even shorter.
minor comments (5)
  1. [Table 4 and Table 5] The variant label 'SPCA Training' should read 'SCPA Training' for consistency with Section 3.2.
  2. [Eq. (7)] There is an unbalanced parenthesis: exp(−λ·sg[E_geo_i])) has one extra closing parenthesis.
  3. [Eq. (6)] The denominator median(D_i) is not defined precisely: is D_i a per-primitive depth value, a per-pixel depth map over a patch, or a scalar? This affects the scale normalization and should be stated.
  4. [Table 4] The row 'Baseline + ROM' does not specify which pose-sampling scheme (original DA3 protocol, context-only, or context-target) is used when ROM is added in isolation. This is needed for an incremental ablation comparison.
  5. [Sec. 4.2] The paper uses the term 'dense-view benchmarks' but the maximum input is 36 views; consider clarifying the sense of 'dense' relative to the sparse-view baseline literature.

Circularity Check

0 steps flagged

No load-bearing circularity: SCPA and ROM are GT-supervised/external-teacher regularizers; the SOTA margins are empirical benchmark results.

full rationale

The claimed derivation chain is not circular. SCPA (Eqs. 1-4) computes a corrected target pose P_align = exp(-log(P2 P1^{-1})) P1 using the model's own pose head on a rendered proxy, then uses L_scpa = min(L_rec(I_align, I_gt), L_rec(I1, I_gt)); although this is self-referential in the sense that the correction is a function of the model's own outputs, the supervision is still ground-truth images and the corrected pose is not a fitted parameter nor a benchmark prediction. ROM (Eqs. 6-8) obtains geometric ratings from a fixed external teacher (DepthSplat [46]) and matches student opacity to those ratings via a margin loss; this is a regularizer with external supervision, not a prediction of the teacher's own outputs. The reported SOTA margins on RE10K/DL3DV/ACID are empirical comparisons against external baselines, and no equation reduces the claimed result to the model's inputs. Two non-circular flags: (1) Sec. 4.2 cites [31] (Ecosplat, same research group) for the evaluation protocol and ACID test splits; this is a minor self-citation but it fixes only the protocol/split, not the derivation, and the benchmarks are external. (2) Sec. 4.2's phrase 'we adopt the test-time pose optimization for ground truth alignment for all pose-free NVS baselines' is ambiguous about whether AirSplat receives the same alignment; this is a fairness/protocol issue that should be clarified, but is not an instance of a prediction reducing to its inputs. Therefore score 1, reflecting the minor self-citation/protocol ambiguity rather than any load-bearing circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim rests on an empirical drift correction, a teacher-rating assumption, and a uniform evaluation protocol. No new physical entities are introduced. All free parameters are validation-set hyperparameters, not fitted to test data.

free parameters (6)
  • λ (rating decay rate) = 5.0
    Hyperparameter governing exponential decay of teacher rating in Eq. 7; empirically chosen.
  • λ_geo = 0.1
    Loss weight for spatial regularization in Eq. 9-10; empirically chosen.
  • λ_opa = 1.0
    Loss weight for opacity matching in Eq. 8 and 10; empirically chosen.
  • λ_s = 0.1
    LPIPS weight in Eq. 5; empirically chosen.
  • τ (spatial error clamp) = 2.0
    Clamp for geometric error in Eq. 9 to avoid exploding gradients; set by hand.
  • η (stability constant) = 1e-7 or 1e-8
    Small constant in Eq. 6 to avoid division by zero; set for numerical stability.
axioms (5)
  • domain assumption The frozen pose head of DA3 exhibits a systematic drift when re-predicting poses from rendered images, so the relative SE(3) transformation between the first and second pose predictions provides a reliable correction direction.
    Empirical observation in Fig. 9; no theoretical guarantee is given.
  • domain assumption The teacher model f̃_ϕ's geometric rating, based on the excess geometric error, is a valid upper bound for geometric consistency, and one-sided opacity matching to this rating prunes artifacts without harming valid semi-transparent primitives.
    Assumes the teacher's normalized geometric error is comparable to the student's and that one-sided matching is sufficient.
  • domain assumption Test-time pose optimization/alignment to ground truth is applied uniformly to all compared pose-free methods, including AirSplat.
    The text in Sec 4.2 says 'for all pose-free NVS baselines' but does not explicitly state AirSplat is included; the SOTA claim depends on this uniformity.
  • domain assumption Freezing the encoder and pose heads while fine-tuning only the Gaussian prediction head preserves the 3DVFM's geometric priors.
    Validated indirectly by Table 6, which shows geometry metrics are unchanged.
  • standard math Standard SE(3) logarithm and exponential map formulations are used correctly.
    Used in Eqs. 2-3 for computing the pose correction.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of AirSplat: Alignment and Rating for Robust Feed-Forward 3D Gaussian Splatting." pith.science (2026). https://pith.science/paper/IAEV7UDM

@misc{pith2026260325129,
  author       = {Pith},
  title        = {Pith review of: AirSplat: Alignment and Rating for Robust Feed-Forward 3D Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IAEV7UDM}},
  note         = {Machine review of arXiv:2603.25129}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

While 3D Vision Foundation Models (3DVFMs) have demonstrated remarkable zero-shot capabilities in visual geometry estimation, their direct application to generalizable novel view synthesis (NVS) remains challenging. In this paper, we propose AirSplat, a novel training framework that effectively adapts the robust geometric priors of 3DVFMs into high-fidelity, pose-free NVS. Our approach introduces two key technical contributions: (1) Self-Consistent Pose Alignment (SCPA), a training-time feedback loop that ensures pixel-aligned supervision to resolve pose-geometry discrepancy; and (2) Rating-based Opacity Matching (ROM), which leverages the local 3D geometry consistency knowledge from a sparse-view NVS teacher model to filter out degraded primitives. Experimental results on large-scale benchmarks demonstrate that our method significantly outperforms state-of-the-art pose-free NVS approaches in reconstruction quality. Our AirSplat highlights the potential of adapting 3DVFMs to enable simultaneous visual geometry estimation and high-quality view synthesis.

Figures

Figures reproduced from arXiv: 2603.25129 by Jaeho Moon, Minh-Quan Viet Bui, Munchurl Kim.

Figure 1
Figure 1. Figure 1: Our proposed AirSplat adapts 3D-VS-VFMs using Self-Consistent Pose Align￾ment (SCPA) and Rating-based Opacity Matching (ROM) to resolve inherent pose￾geometry discrepancies and multi-view inconsistencies. Our AirSplat effectively elimi￾nates ‘floaters’ (red boxes) and blurry artifacts (dashed yellow boxes) produced by the baseline DA3 [20], rendering sharp, structurally consistent novel views. Abstract. Wh… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our AirSplat training pipeline. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of training paradigms in pose-free NVS. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect of Self-Consistent Pose Alignment (SCPA). [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of NVS performance on RE10K dataset [54]. 4.3 NVS Performance Evaluation Comparison on RE10K. We evaluate our model on the RE10K dataset to verify its scalability and robustness in large-scale indoor and outdoor environ￾ments. As shown in [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison of NVS performance on DL3DV dataset [21]. Comparison on DL3DV [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of the ‘floaters’ compression in the predicted 3DGS per context views [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Our strategy for partitioning input context sequences to compute the teacher’s geometric errors. GT 𝑰𝑰� tgt,t align 𝑰𝑰� tgt,t (1) 𝑰𝑰� tgt,t (2) 𝑰𝑰� tgt,t (3) (a) (b) 19.92dB 15.69dB 14.10dB 12.96dB 29.03dB 19.71dB 16.89dB 15.91dB [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Pose-Geometry Discrepancy. To validate that this structural misalignment is not confined to a specific baseline, we visualize the consistent spatial drift across distinct models and diverse scenes: (a) DepthAnything3 [20] evaluated on the DL3DV dataset [21], and (b) SPFSplat [13] on the RE10K dataset [54]. dataset [21], alongside the discrepancy found in SPFSplat [13] on the RE10K dataset [54]. In the figu… view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative comparison of NVS performance on DL3DV dataset [21]. rendered novel views. To resolve these deterministic blind spots, future direc￾tions of this framework could integrate generative video diffusion priors as a post-processing module, allowing for the temporally consistent inpainting of ge￾ometrically plausible content within these occlusions. H Additional Qualitative Comparison [PITH_FULL_IM… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

55 extracted references · 7 linked inside Pith

  1. [1]

    arXiv preprint arXiv:1902.04257 (2019)

    Arumugam, D., Lee, J.K., Saskin, S., Littman, M.L.: Deep reinforcement learning from policy-dependent human feedback. arXiv preprint arXiv:1902.04257 (2019)

    Pith/arXiv arXiv 1902

  2. [2]

    IEEE Access (2024)

    Bello, J.L.G., Bui, M.Q.V., Kim, M.: Pronerf: Learning efficient projection-aware ray sampling for fine-grained implicit neural radiance fields. IEEE Access (2024). https://doi.org/10.1109/ACCESS.2024.3390753

    arXiv 2024

  3. [3]

    CoRR (2021)

    Blanco-Claraco, J.L.: A tutorial on $\mathbf{SE}(3)$ transformation parameter- izations and on-manifold optimization. CoRR (2021)

    2021

  4. [4]

    In: AAAI (2026)

    Bui, M.Q.V., Park, J., Bello, J.L.G., Moon, J., Oh, J., Kim, M.: Mobgs: Motion deblurring dynamic 3d gaussian splatting for blurry monocular video. In: AAAI (2026)

    2026

  5. [5]

    In: CVPR (2024)

    Charatan,D.,Li,S.L.,Tagliasacchi,A.,Sitzmann,V.:pixelsplat:3dgaussiansplats from image pairs for scalable generalizable 3d reconstruction. In: CVPR (2024)

    2024

  6. [6]

    In: ECCV (2022)

    Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: ECCV (2022)

    2022

  7. [7]

    In: ICCV (2021)

    Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: ICCV (2021)

    2021

  8. [8]

    In: ECCV (2024)

    Chen, Y., Xu, H., Zheng, C., Zhuang, B., Pollefeys, M., Geiger, A., Cham, T.J., Cai, J.: Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In: ECCV (2024)

    2024

  9. [9]

    In: CVPR (2022)

    Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenox- els: Radiance fields without neural networks. In: CVPR (2022)

    2022

  10. [10]

    In: Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition

    Fu, Y., Liu, S., Kulkarni, A., Kautz, J., Efros, A.A., Wang, X.: Colmap-free 3d gaussian splatting. In: Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition. pp. 20796–20805 (2024)

    2024

  11. [11]

    In: CVPR (2024)

    Hong, S., Jung, J., Shin, H., Yang, J., Kim, S., Luo, C.: Coponerf: Unifying corre- spondence, pose and nerf for pose-free novel view synthesis from stereo pairs. In: CVPR (2024)

    2024

  12. [12]

    In: ACM SIGGRAPH 2024 conference papers

    Huang, B., Yu, Z., Chen, A., Geiger, A., Gao, S.: 2d gaussian splatting for geo- metrically accurate radiance fields. In: ACM SIGGRAPH 2024 conference papers. pp. 1–11 (2024)

    2024

  13. [13]

    In: ICCV (2025)

    Huang, R., Mikolajczyk, K.: No pose at all: Self-supervised pose-free 3d gaussian splatting from sparse views. In: ICCV (2025)

    2025

  14. [14]

    In: ICCV (2025)

    Jiang, H., Tan, H., Wang, P., Jin, H., Zhao, Y., Bi, S., Zhang, K., Luan, F., Sunkavalli, K., Huang, Q., Pavlakos, G.: Rayzer: A self-supervised large view syn- thesis model. In: ICCV (2025)

    2025

  15. [15]

    arXiv preprint arXiv:2505.23716 (2025)

    Jiang, L., Mao, Y., Xu, L., Lu, T., Ren, K., Jin, Y., Xu, X., Yu, M., Pang, J., Zhao, F., et al.: Anysplat: Feed-forward 3d gaussian splatting from unconstrained views. arXiv preprint arXiv:2505.23716 (2025)

    arXiv 2025

  16. [16]

    In: CVPR (2025)

    Kang, G., Yoo, J., Park, J., Nam, S., Im, H., Shin, S., Kim, S., Park, E.: Selfsplat: Pose-free and 3d prior-free generalizable 3d gaussian splatting. In: CVPR (2025)

    2025

  17. [17]

    ACM Trans

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. (2023)

    2023

  18. [18]

    In: ECCV (2024)

    Leroy, V., Cabon, Y., Revaud, J.: Grounding image matching in 3d with mast3r. In: ECCV (2024)

    2024

  19. [19]

    In: ICCV (2025)

    Lin,C.Y.,Sun,C.,Yang,F.E.,Chen,M.H.,Lin,Y.Y.,Liu,Y.L.:Longsplat:Robust unposed 3d gaussian splatting for casual long videos. In: ICCV (2025)

    2025

  20. [20]

    arXiv preprint arXiv:2511.10647 (2025) AirSplat 7

    Lin, H., Chen, S., Liew, J., Chen, D.Y., Li, Z., Shi, G., Feng, J., Kang, B.: Depth anything 3: Recovering the visual space from any views. arXiv preprint arXiv:2511.10647 (2025) AirSplat 7

    Pith/arXiv arXiv 2025

  21. [21]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Ling, L., Sheng, Y., Tu, Z., Zhao, W., Xin, C., Wan, K., Yu, L., Guo, Q., Yu, Z., Lu, Y., et al.: Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22160–22169 (2024)

    2024

  22. [22]

    In: ICCV (2021)

    Liu, A., Tucker, R., Jampani, V., Makadia, A., Snavely, N., Kanazawa, A.: Infinite nature: Perpetual view generation of natural scenes from a single image. In: ICCV (2021)

    2021

  23. [23]

    In: CVPR (2025)

    Liu, Y., Fan, K., Yu, W., Li, C., Lu, H., Yuan, Y.: Monosplat: Generalizable 3d gaussian splatting from monocular depth foundation models. In: CVPR (2025)

    2025

  24. [24]

    Liu, Y., Min, Z., Wang, Z., Wu, J., Wang, T., Yuan, Y., Luo, Y., Guo, C.: World- mirror:Universal3dworldreconstructionwithany-priorprompting.arXivpreprint arXiv:2510.10726 (2025)

    Pith/arXiv arXiv 2025

  25. [25]

    arXiv preprint arXiv:1711.05101 (2017)

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)

    Pith/arXiv arXiv 2017

  26. [26]

    In: ICML (2025)

    Luu,T.M.,Lee,Y.,Lee,D.,Kim,S.,Kim,M.J.,Yoo,C.D.:Enhancingrating-based reinforcement learning to effectively leverage feedback from large vision-language models. In: ICML (2025)

    2025

  27. [27]

    In: ICML (2017)

    MacGlashan, J., Ho, M.K., Loftin, R., Peng, B., Wang, G., Roberts, D.L., Taylor, M.E., Littman, M.L.: Interactive learning from policy-dependent human feedback. In: ICML (2017)

    2017

  28. [28]

    In: ECCV (2020)

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)

    2020

  29. [29]

    ACM transactions on graphics (TOG)41(4), 1–15 (2022)

    Müller,T.,Evans,A.,Schied,C.,Keller,A.:Instantneuralgraphicsprimitiveswith a multiresolution hash encoding. ACM transactions on graphics (TOG)41(4), 1–15 (2022)

    2022

  30. [30]

    In: CVPR (2025)

    Park, J., Bui, M.Q.V., Bello, J.L.G., Moon, J., Oh, J., Kim, M.: Splinegs: Robust motion-adaptive spline for real-time dynamic 3d gaussians from monocular video. In: CVPR (2025)

    2025

  31. [31]

    In: CVPR (2026)

    Park, J., Bui, M.Q.V., Bello, J.L.G., Moon, J., Oh, J., Kim, M.: Ecosplat: Efficiency-controllable feed-forward 3d gaussian splatting from multi-view images. In: CVPR (2026)

    2026

  32. [32]

    In: CVPR (2016)

    Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016)

    2016

  33. [33]

    In: CVPR (2013)

    Shotton, J., Glocker, B., Zach, C., Izadi, S., Criminisi, A., Fitzgibbon, A.: Scene coordinate regression forests for camera relocalization in rgb-d images. In: CVPR (2013)

    2013

  34. [34]

    arXiv preprint arXiv:2408.13912 (2024)

    Smart, B., Zheng, C., Laina, I., Prisacariu, V.A.: Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912 (2024)

    Pith/arXiv arXiv 2024

  35. [35]

    In: Artificial intelligence and machine learning for multi- domain operations applications

    Smith, L.N., Topin, N.: Super-convergence: Very fast training of neural networks using large learning rates. In: Artificial intelligence and machine learning for multi- domain operations applications. vol. 11006, pp. 369–386. SPIE (2019)

    2019

  36. [36]

    ACM Trans

    Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3d. ACM Trans. Graph. (2006)

    2006

  37. [37]

    In: ECCV (2022)

    Suhail, M., Esteves, C., Sigal, L., Makadia, A.: Generalizable patch-based neural rendering. In: ECCV (2022)

    2022

  38. [38]

    In: CVPR (2024)

    Szymanowicz, S., Rupprecht, C., Vedaldi, A.: Splatter image: Ultra-fast single-view 3d reconstruction. In: CVPR (2024)

    2024

  39. [39]

    In: CVPR (2025) 8 Bui et al

    Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotny, D.: Vggt: Visual geometry grounded transformer. In: CVPR (2025) 8 Bui et al

    2025

  40. [40]

    In: CVPR (2021)

    Wang, Q., Wang, Z., Genova, K., Srinivasan, P., Zhou, H., Barron, J.T., Martin- Brualla, R., Snavely, N., Funkhouser, T.: Ibrnet: Learning multi-view image-based rendering. In: CVPR (2021)

    2021

  41. [41]

    In: CVPR (2024)

    Wang, S., Leroy, V., Cabon, Y., Chidlovskii, B., Revaud, J.: Dust3r: Geometric 3d vision made easy. In: CVPR (2024)

    2024

  42. [42]

    arXiv preprint arXiv:2505.23734 (2025)

    Wang, W., Chen, D.Y., Zhang, Z., Shi, D., Liu, A., Zhuang, B.: Zpressor: Bottleneck-aware compression for scalable feed-forward 3dgs. arXiv preprint arXiv:2505.23734 (2025)

    arXiv 2025

  43. [43]

    arXiv preprint arXiv:2507.13347 (2025)

    Wang, Y., Zhou, J., Zhu, H., Chang, W., Zhou, Y., Li, Z., Chen, J., Pang, J., Shen, C., He, T.:π 3: Permutation-equivariant visual geometry learning. arXiv preprint arXiv:2507.13347 (2025)

    Pith/arXiv arXiv 2025

  44. [44]

    In: AAAI (2024)

    White, D., Wu, M., Novoseller, E., Lawhern, V.J., Waytowich, N., Cao, Y.: Rating- based reinforcement learning. In: AAAI (2024)

    2024

  45. [45]

    In: CVPR (2024)

    Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. In: CVPR (2024)

    2024

  46. [46]

    In: CVPR (2025)

    Xu, H., Peng, S., Wang, F., Blum, H., Barath, D., Geiger, A., Pollefeys, M.: Depth- splat: Connecting gaussian splatting and depth. In: CVPR (2025)

    2025

  47. [47]

    In: ICLR (2026)

    Ye, B., Chen, B., Xu, H., Barath, D., Pollefeys, M.: Yonosplat: You only need one model for feedforward 3d gaussian splatting. In: ICLR (2026)

    2026

  48. [48]

    arXiv preprint arXiv:2410.24207 (2024)

    Ye, B., Liu, S., Xu, H., Li, X., Pollefeys, M., Yang, M.H., Peng, S.: No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207 (2024)

    Pith/arXiv arXiv 2024

  49. [49]

    In: ICCV (2021)

    Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for real-time rendering of neural radiance fields. In: ICCV (2021)

    2021

  50. [50]

    In: CVPR (2021)

    Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: CVPR (2021)

    2021

  51. [51]

    In: Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition

    Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting. In: Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition. pp. 19447–19456 (2024)

    2024

  52. [52]

    In: CVPR (2018)

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)

    2018

  53. [53]

    In: CVPR (2025)

    Zhang, S., Wang, J., Xu, Y., Xue, N., Rupprecht, C., Zhou, X., Shen, Y., Wet- zstein, G.: Flare: Feed-forward geometry, appearance and camera estimation from uncalibrated sparse views. In: CVPR (2025)

    2025

  54. [54]

    In: SIGGRAPH (2018)

    Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo magnification: Learn- ing view synthesis using multiplane images. In: SIGGRAPH (2018)

    2018

  55. [55]

    In: ICCV (2025)

    Ziwen, C., Tan, H., Zhang, K., Bi, S., Luan, F., Hong, Y., Fuxin, L., Xu, Z.: Long- lrm: Long-sequence large reconstruction model for wide-coverage gaussian splats. In: ICCV (2025)

    2025

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.