Pith. sign in

REVIEW 3 major objections 4 minor 69 references

UFV-Splatter: Pose-Free Feed-Forward 3D Gaussian Splatting Adapted to Unfavorable Views

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a pretrained pose-free 3D Gaussian splatting model, adapted with recentering, LoRA layers, and a Gaussian adapter trained on teacher-rendered unfavorable views, can synthesize consistent novel views from off-center…

desk verdict A useful adaptation method with a real evaluation confound: per-scene alignment is applied to all methods, and the teacher-distillation training loop is disclosed but deserves harder testing. read the letter →

arxiv 2507.22342 v2 pith:36UTOX3C submitted 2025-07-30 cs.CV

Yuki Fujimura , Takahiro Kushida , Kazuya Kitano , Takuya Funatomi , Yasuhiro Mukaigawa This is my paper
classification cs.CV
keywords pose-free3DGaussiansplattingunfavorableviewsnovelviewsynthesislow-rankadaptationadaptermodelrecenteringsparse
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Pose-free feed-forward 3D Gaussian splatting models are trained almost exclusively on favorable views, with the object centered at the world origin and cameras pointed at it. This paper claims that such a pretrained model can be adapted to unfavorable views — off-center, unknown camera poses, as in handheld photography — without collecting any new unfavorable-view data. The adaptation works by recentering each input so the pretrained model's favorable-view priors still apply, inserting LoRA layers into the backbone, adding a transformer adapter that repairs the geometry the recentering distorts, and aligning predicted Gaussians to teacher-rendered point maps during training. On synthetic Google Scanned Objects images the adapted model reports PSNR 22.910 versus 16.698 for the recentered baseline, and on real OmniObject3D images 18.861 versus 15.807. If the claim holds, pose-free 3D reconstruction could move from constrained turntable-style captures to everyday, unknown-pose photos.

What carries the argument

The load-bearing mechanism is the recentering-plus-alignment loop. Recentering shifts and resizes the foreground bounding box to a canonical size, so the pretrained model sees inputs resembling its favorable-view training data; the recentered 2D positional embeddings preserve where each patch came from, which the adapter uses to correct geometry. The Gaussian adapter is a set of self-attention blocks that takes predicted per-pixel Gaussians, adds the recentered positional embedding, and outputs residuals that are added to the original Gaussians. Gaussian alignment solves a weighted least-squares problem for a global scale and translation that maps predicted Gaussian centers onto teacher-rendered point maps, rendering accurate target views for the training loss. LoRA layers inserted into the patchify convolution and self-attention blocks provide the parameter-efficient adaptation that lets the pretrained model absorb the new input distribution.

What would settle it

Render the same G-Objaverse samples in two ways — once from the frozen teacher's Gaussians and once from ground-truth meshes with identical unfavorable camera translations — train the identical pipeline on each, and compare on the real OmniObject3D test set. If the mesh-rendered training data gives materially higher PSNR than the teacher-rendered data, the teacher's favorable-view Gaussians are the bottleneck and the transfer claim is only as strong as the teacher.

Watch

Extended reading notes

Core claim

The central claim is that a pose-free feed-forward 3DGS model pretrained on favorable views can be made to handle unfavorable views by a combination of input recentering, low-rank adaptation, a Gaussian adapter, and teacher-generated training images. Recentering maps each off-center capture into the distribution the pretrained model saw, so its learned priors remain usable. LoRA layers let the backbone adapt cheaply to the recentered inputs. The adapter, a small cross-view transformer operating on Gaussian tokens plus recentered positional embeddings, predicts residual corrections to the Gaussians, repairing the pixel-to-ray distortion introduced by recentering. During training, the frozen pretrained model estimates Gaussians from favorable images of an off-the-shelf dataset; those Gaussians are rendered from translated, random-FoV cameras to synthesize unfavorable context views; and a closed-form weighted least-squares alignment rescales and shifts the student's predicted Gaussians so rendered targets are correctly supervised. The paper reports that this pipeline outperforms pose-free baselines on both synthetic and real object datasets.

Load-bearing premise

The paper assumes the frozen pretrained model's Gaussians, learned from favorable views, render unfavorable context views faithfully enough that training on those renderings transfers to real off-center captures.

Editorial extensions

If this is right

  • Pose-free feed-forward 3DGS can be adapted to a new viewpoint distribution using only existing favorable-view datasets, removing the need to build large multi-view captures with unfavorable poses.
  • The recentering step makes the method compatible with everyday input where the object is not centered and the camera pose is unknown, as long as the foreground can be segmented.
  • Adding more context views degrades unadapted baselines while the adapted model stays stable, so the method scales to larger sparse-view sets.
  • On favorable views the adapted model keeps most of the original quality (PSNR 23.332 recentered versus 24.082 unadapted), so a single adapted model can serve unknown view conditions.
  • Because inputs are recentered before inference, the adapted model gives up direct camera-pose recovery via PnP, a trade-off the paper states explicitly.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The training data are generated by the frozen teacher itself, so the method's ceiling is set by teacher quality: a stronger pose-free teacher should transfer directly to higher unfavorable-view accuracy, which is testable by swapping the backbone.
  • Because the alignment used in training is a closed-form similarity fit to point maps, the same alignment could be applied at inference time to canonicalize predicted Gaussians, potentially removing the need to render teacher targets during training.
  • The real-data gain on OmniObject3D is smaller than the synthetic gain on GSO, which is consistent with a simulation gap between teacher-rendered unfavorable views and true off-center captures; evaluating on real images with known poses would isolate that gap.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes UFV-Splatter, an adaptation of the pretrained pose-free feed-forward 3D Gaussian Splatting model FreeSplatter to handle unfavorable input views, defined as off-center, unknown-pose cameras. The method recenters input images and positional embeddings, inserts LoRA layers into the backbone, adds a small transformer-based Gaussian adapter that predicts residual corrections, and uses a closed-form weighted-least-squares alignment of predicted Gaussians to render supervision targets. Training uses only a favorable-view dataset (G-Objaverse) by first estimating teacher Gaussians with the frozen FreeSplatter and then rendering unfavorable context views and target views from those Gaussians. Experiments on Google Scanned Objects and OmniObject3D report large improvements over FreeSplatter and LucidFusion, with ablations in Table 2 and supplementary material.

Significance. If valid, the paper would provide a practical recipe for extending pose-free feed-forward 3DGS to real-world, off-center, unknown-pose inputs without collecting new unfavorable-view training data. The strengths are the clearly derived closed-form alignment (Sec. 3.4 and Sec. 8 of the supplementary), the component ablations in Table 2, and the inclusion of a real-image benchmark (OmniObject3D). However, the two main load-bearing elements -- the teacher-generated training supervision and the evaluation protocol that applies per-scene alignment to all methods -- currently leave the headline claim only partially supported. The work is a reasonable incremental contribution in a competitive area, but the evidence needs to be strengthened before the central claim can be accepted.

major comments (3)
  1. [Sec. 3.4, Fig. 4, Eq. (4)] The training supervision is generated by the frozen pretrained FreeSplatter itself: favorable G-Objaverse images are fed to the teacher, and the teacher's Gaussians are used both to render the 'unfavorable' context views and to supply the target point maps P_c^i in Eq. (4). The loss is computed against renderings of the teacher, not against real observations at those viewpoints. Consequently, the student is optimized to reproduce the teacher's reconstruction behavior at off-center poses, including any artifacts. This is load-bearing because the paper claims adaptation without any unfavorable-view dataset. The smaller real-data gain on OmniObject3D (18.861 vs 15.807, about 3.1 dB) compared with GSO (22.910 vs 16.698, about 6.2 dB) is consistent with a simulation-to-real gap. Please add a quantitative analysis of the teacher's rendering quality on held-out favorable views, or a variant trained with real unfavorable-view data, to demonstrate that the adaptation transfers beyond the teacher's distribution.
  2. [Sec. 4.1, Table 1, Fig. 6 caption] All methods except LEAP are evaluated with the proposed Gaussian alignment applied, using per-scene depth maps: rendered depth for GSO and 2DGS-estimated depth for OmniObject3D. The reported metrics therefore measure the feed-forward prediction plus a scene-specific similarity transform fitted to target-scene geometry, not the model's pose-free generalization in isolation. This is especially problematic because the alignment is part of the proposed training pipeline, so the evaluation is not an apples-to-apples comparison of feed-forward models. Please report results without alignment, or with a fixed/common alignment, and clarify how the alignment would be obtained in a real deployment where target-scene depth is not available.
  3. [Sec. 4.2, Table 1] The comparison is affected by a resolution mismatch: the proposed method and FreeSplatter are evaluated at 512x512, while LucidFusion is resized to 256x256 and LEAP to 224x224. If the lower-resolution outputs are compared directly or naively upsampled to the 512x512 ground truth, the reported gains over these baselines are inflated. Please state the resolution at which the metrics in Table 1 are computed and, ideally, provide a resolution-matched comparison for the baselines.
minor comments (4)
  1. [Table 3] The header 'PNSR' is a typo and should read 'PSNR'.
  2. [Table 2] The checkmark layout in the ablation table is ambiguous: it is not immediately clear which components are active in each row. Please make the row labels explicit (e.g., 'LoRA only', 'LoRA + Adapter', 'LoRA + Adapter + PE').
  3. [Sec. 4.1] The description of the OmniObject3D evaluation should clarify whether the 'same four sparse views used in [51]' are already unfavorable, or whether translations are applied to make them unfavorable; the definition of unfavorable views in Sec. 3.2 is given for synthetic data only.
  4. [Figures 6 and 7] Several figure captions are incomplete sentences; for example, Figure 6 has only 'Results of novel-view synthesis on the GSO and OmniObject3D dataset' and Figure 7 does not explain the overlay markers. Please expand the captions so that they are self-contained.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the training labels are teacher-generated, but the central claim is validated on external ground truth and baselines receive identical alignment.

full rationale

The strongest candidate for circularity is the Sec. 3.4 training pipeline, where favorable G-Objaverse images are fed to the frozen pretrained FreeSplatter, and the resulting teacher Gaussians are used both to render the 'unfavorable' context/target views and to provide the point-map alignment targets in Eq. (4). This is a disclosed self-distillation design: the student is trained to reproduce renderings of the teacher's own Gaussians, so any teacher artifacts at off-center poses could in principle be learned. However, this is a simulation-gap risk, not a circular derivation, because the paper's central claim is not forced by construction. The evaluation in Sec. 4.3, Table 1 is against external ground truth: GSO test views are rendered from real 3D scans and OmniObject3D test views are real captures, and the same Gaussian-alignment post-processing is applied to FreeSplatter w/ and w/o recentering and to LucidFusion. The reported gains (22.910 vs 16.698 PSNR on GSO; 18.861 vs 15.807 on OmniObject3D) therefore cannot be reduced to an identity with the teacher's outputs or to the fitted alignment. There is no load-bearing self-citation (the authors are not the FreeSplatter authors), no imported uniqueness theorem, and no ansatz smuggled in via citation. The paper itself acknowledges the recentering-based limitation and the object-centric assumption. Under the rubric's strict requirement that circularity be exhibited by a specific reduction equation or a fitted parameter renamed as a prediction, the teacher-generated supervision is a validity concern but not a circular step, and the derivation remains independently checkable against external benchmarks.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on one external teacher (FreeSplatter), two domain assumptions about simulation fidelity (teacher rendering and 2DGS depth), and seven hand-chosen hyperparameters. The per-scene alignment absorbs the scale/position ambiguity that the model never learns to resolve, which is the single most consequential fitted quantity. No new physical or architectural entities are postulated beyond the disclosed adapter module.

free parameters (7)
  • Per-scene alignment scale a and translation b = fitted per scene via weighted least squares (Eq. 4)
    Solved in closed form against depth-derived point maps during training and evaluation; they absorb the global scale and position ambiguity of the pose-free Gaussians, so the model itself never learns to resolve that ambiguity.
  • Recentering resize factor 0.8 = 0.8
    Chosen by hand; foreground is resized so max(Hf,Wf) = 0.8 min(H,W) (Sec 3.3).
  • Unfavorable-view translation magnitude rho = uniform in [-0.1r, 0.1r] per axis for training; 0.1 and 0.2 shown in Fig 2
    Chosen by hand; defines the distribution of camera offset from the object center.
  • LoRA rank = 64
    Hyperparameter, adopted after the appendix comparison of insertion configurations.
  • LPIPS loss weight = 0.5
    MSE and LPIPS weighted by 0.5 in all experiments (Sec 3.4).
  • Gaussian center residual bound = atanh scaled by 0.05 per axis
    Adapter residuals are constrained to [-0.05, 0.05] to stay within the rasterizer's expected parameter range (supplement Sec 6.2).
  • Camera sampling ranges for training = FoV [20,50] deg, elevation [0,45] deg, azimuth offsets 90 deg
    Chosen by hand; defines the rendered training views (supplement Sec 7).
assumptions (5)
  • domain assumption The frozen pretrained FreeSplatter produces accurate enough Gaussians from favorable views to serve as training labels.
    The entire self-training pipeline in Sec 3.4 renders the student's training views from teacher Gaussians; if the teacher's geometry is wrong at novel viewpoints, the student learns those errors.
  • domain assumption Unfavorable views rendered from teacher Gaussians faithfully simulate real off-center captures in occlusion, background, and lighting.
    Sec 3.4 uses these synthetic views as the only training signal for the unfavorable-view distribution; the smaller OmniObject3D gains suggest the simulation only partially transfers.
  • domain assumption Depth maps estimated by 2D Gaussian Splatting on 200 frames are accurate enough for the alignment during evaluation.
    Sec 4.1 applies 2DGS depth for alignment of every method; no error analysis of these depth maps is provided, and alignment quality bounds the reported numbers.
  • domain assumption Object-centric scenes with well-segmented foregrounds.
    Recentering requires a foreground mask (mask [61]); the paper states this as a limitation in Sec 4.4.
  • standard math Weighted least-squares closed form and transformer self-attention behave as standard.
    Eq. 4 and supplement Eqs. 12-13 use vanilla linear algebra; transformer blocks are standard.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UFV-Splatter: Pose-Free Feed-Forward 3D Gaussian Splatting Adapted to Unfavorable Views." pith.science (2026). https://pith.science/paper/36UTOX3C

@misc{pith2026250722342,
  author       = {Pith},
  title        = {Pith review of: UFV-Splatter: Pose-Free Feed-Forward 3D Gaussian Splatting Adapted to Unfavorable Views},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/36UTOX3C}},
  note         = {Machine review of arXiv:2507.22342}
}
read the original abstract

This paper presents a pose-free, feed-forward 3D Gaussian Splatting (3DGS) framework designed to handle unfavorable input views. A common rendering setup for training feed-forward approaches places a 3D object at the world origin and renders it from cameras pointed toward the origin -- i.e., from favorable views, limiting the applicability of these models to real-world scenarios involving varying and unknown camera poses. To overcome this limitation, we introduce a novel adaptation framework that enables pretrained pose-free feed-forward 3DGS models to handle unfavorable views. We leverage priors learned from favorable images by feeding recentered images into a pretrained model augmented with low-rank adaptation (LoRA) layers. We further propose a Gaussian adapter module to enhance the geometric consistency of the Gaussians derived from the recentered inputs, along with a Gaussian alignment method to render accurate target views for training. Additionally, we introduce a new training strategy that utilizes an off-the-shelf dataset composed solely of favorable images. Experimental results on both synthetic images from the Google Scanned Objects dataset and real images from the OmniObject3D dataset validate the effectiveness of our method in handling unfavorable input views.

Figures

Figures reproduced from arXiv: 2507.22342 by the authors.

Figure 1
Figure 1. Novel-view synthesis for daily captures with a smart [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Favorable views (red cameras, ρ = 0): the object is placed at the world origin, and the cameras are oriented toward the origin. Unfavorable views (green and blue cameras, ρ = 0.1, 0.2): generated by adding a translation of random direction and magnitude ρr to each favorable camera, where r denotes the distance from the object origin to the camera. FreeSplatter [50], trained only on favorable views, performs well whe… view at source ↗
Figure 3
Figure 3. Overview of the proposed model. It is built upon a pretrained pose-free feed-forward 3DGS model, augmented with LoRA [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Overview of the training framework without requiring an unfavorable-view dataset. Favorable images from an off-the-shelf [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Effect of Gaussian alignment during training. (a) [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Results of novel-view synthesis on the GSO and OmniObject3D dataset [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 8
Figure 8. Figure 8: Performance comparison under different numbers of [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Examples of rendered images used for training. (a) Favorable target images. (b) Unfavorable context images [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Additional results of novel-view synthesis on the GSO dataset [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Additional results of novel-view synthesis on the OmniObject3D dataset [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

69 extracted references · 62 canonical work pages

  1. [1]

    pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction

    David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In CVPR, pages 19457–19467, 2024. 1, 2

    work page 2024

  2. [2]

    Mvsnerf: Fast general- izable radiance field reconstruction from multi-view stereo

    Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast general- izable radiance field reconstruction from multi-view stereo. In ICCV, pages 14124–14133, 2021. 1, 2

    work page 2021

  3. [3]

    Lara: Efficient large-baseline radiance fields

    Anpei Chen, Haofei Xu, Stefano Esposito, Siyu Tang, and Andreas Geiger. Lara: Efficient large-baseline radiance fields. In ECCV, pages 338–355, 2024. 2, 6

    work page 2024

  4. [4]

    Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images

    Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In ECCV, pages 370–386, 2024. 1, 2

    work page 2024

  5. [5]

    Mvsplat360: Feed-forward 360 scene synthesis from sparse views

    Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views. In NeurIPS, 2024. 1, 2

    work page 2024

  6. [6]

    Depth-regularized optimization for 3d gaussian splatting in few-shot images

    Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. In CVPRW, pages 811–820, 2024. 2

    work page 2024

  7. [7]

    Objaverse: A universe of annotated 3d objects

    Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. In CVPR, pages 13142– 13153, 2023. 1, 6

    work page 2023

  8. [8]

    Depth-supervised nerf: Fewer views and faster train- ing for free

    Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ra- manan. Depth-supervised nerf: Fewer views and faster train- ing for free. In CVPR, pages 12882–12891, 2022. 2

    work page 2022

Show all 69 references
  1. [9]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition ...

    2021

  2. [10]

    McHugh, and Vincent Vanhoucke

    Laura Downs, Anthony Francis, Nate Koenig, Brandon Kin- man, Ryan Hickman, Krista Reymann, Thomas B. McHugh, and Vincent Vanhoucke. Google scanned objects: A high- quality dataset of 3d scanned household items. InICRA, page 2553–2560, 2022. 6

    2022

  3. [11]

    Instantsplat: Sparse-view sfm-free gaussian splatting in seconds

    Zhiwen Fan, Kairun Wen, Wenyan Cong, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, Zhangyang Wang, and Yue Wang. Instantsplat: Sparse-view sfm-free gaussian splatting in seconds. In arXiv preprint arXiv:2403.20309, 2024. 2

    2024 arXiv

  4. [12]

    Sparsenerf: Distilling depth ranking for few-shot novel view synthesis

    Guangcong, Zhaoxi Chen, Chen Change Loy, and Ziwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In ICCV, 2023. 1, 2

    2023

  5. [13]

    Multiple View Ge- ometry in Computer Vision

    Richard Hartley and Andrew Zisserman. Multiple View Ge- ometry in Computer Vision . Cambridge University Press,

  6. [14]

    Lu- cidfusion: Reconstructing 3d gaussians with arbitrary un- posed images

    Hao He, Yixun Liang, Luozhou Wang, Yuanhao Cai, Xinli Xu, Hao-Xiang Guo, Xiang Wen, and Yingcong Chen. Lu- cidfusion: Reconstructing 3d gaussians with arbitrary un- posed images. In arXiv preprint arXiv:2410.15636v3, 2025. 3, 6, 2

    2025 arXiv

  7. [15]

    Pf3plat: Pose-free feed-forward 3d gaussian splatting

    Sunghwan Hong, Jaewoo Jung, Heeseong Shin, Jisang Han, Jiaolong Yang, Chong Luo, and Seungryong Kim. Pf3plat: Pose-free feed-forward 3d gaussian splatting. In ICML,

  8. [16]

    Lrm: Large reconstruction model for single image to 3d

    Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In ICLR, 2024. 2, 3

    2024

  9. [17]

    LoRA: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 2, 4, 1

    2022

  10. [18]

    2d gaussian splatting for geometrically ac- curate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically ac- curate radiance fields. In ACM SIGGRAPH, 2024. 6

    2024

  11. [19]

    Putting nerf on a diet: Semantically consistent few-shot view synthesis

    Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In ICCV, pages 5885–5894, 2021. 1

    2021

  12. [20]

    Few-view object reconstruction with unknown cate- gories and camera poses

    Hanwen Jiang, Zhenyu Jiang, Kristen Grauman, and Yuke Zhu. Few-view object reconstruction with unknown cate- gories and camera poses. In 3DV, 2024. 3, 6

    2024

  13. [21]

    Leap: Liberate sparse-view 3d modeling from camera poses

    Hanwen Jiang, Zhenyu Jiang, Yue Zhao, and Qixing Huang. Leap: Liberate sparse-view 3d modeling from camera poses. In ICLR, 2024. 3, 6

    2024

  14. [22]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 42(4), 2023. 1, 2

    2023

  15. [23]

    Infonerf: Ray entropy minimization for few-shot neural volume ren- dering

    Mijeong Kim, Seonguk Seo, and Bohyung Han. Infonerf: Ray entropy minimization for few-shot neural volume ren- dering. In CVPR, pages 12912–12921, 2022. 2

    2022

  16. [24]

    Genera- tive sparse-view gaussian splatting

    Hanyang Kong, Xingyi Yang, and Xinchao Wang. Genera- tive sparse-view gaussian splatting. In CVPR, pages 26745– 26755, 2025. 1, 2

    2025

  17. [25]

    Gecon- erf: few-shot neural radiance fields via geometric consis- tency

    Min-Seop Kwak, Jiuhn Song, and Seungryong Kim. Gecon- erf: few-shot neural radiance fields via geometric consis- tency. In ICML, 2023. 2

    2023

  18. [26]

    Corresnerf: Image correspondence pri- ors for neural radiance fields

    Yixing Lao, Xiaogang Xu, Zhipeng Cai, Xihui Liu, and Hengshuang Zhao. Corresnerf: Image correspondence pri- ors for neural radiance fields. In NeurIPS, 2023. 2

    2023

  19. [27]

    Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normaliza- tion

    Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normaliza- tion. In CVPR, pages 20775–20785, 2024. 1, 2

    2024

  20. [28]

    Frugalnerf: Fast con- vergence for extreme few-shot novel view synthesis without learned priors

    Chin-Yang Lin, Chung-Ho Wu, Chang-Han Yeh, Shih-Han Yen, Cheng Sun, and Yu-Lun Liu. Frugalnerf: Fast con- vergence for extreme few-shot novel view synthesis without learned priors. In CVPR, pages 11227–11238, 2025. 2

    2025

  21. [29]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6 9

    2019

  22. [30]

    Tracknerf: Bundle adjusting nerf from sparse and noisy views via feature tracks

    Jinjie Mai, Wenxuan Zhu, Sara Rojas, Jesus Zarzar, Abdul- lah Hamdi, Guocheng Qian, Bing Li, Silvio Giancola, and Bernard Ghanem. Tracknerf: Bundle adjusting nerf from sparse and noisy views via feature tracks. In ECCV, pages 470–489, 2024. 2

    2024

  23. [31]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. In ECCV, 2020. 2

    2020

  24. [32]

    Upfusion: Novel view diffusion from unposed sparse view observations

    Bharath Raj Nagoor Kani, Hsin-Ying Lee, Sergey Tulyakov, and Shubham Tulsiani. Upfusion: Novel view diffusion from unposed sparse view observations. InECCV, pages 179–195,

  25. [33]

    Barron, Ben Mildenhall, Mehdi S

    Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view syn- thesis from sparse inputs. In CVPR, pages 5480–5490, 2022. 2

    2022

  26. [34]

    Barron, Ben Mildenhall, Pratul P

    Barbara Roessle, Jonathan T. Barron, Ben Mildenhall, Pratul P. Srinivasan, and Matthias Nießner. Dense depth pri- ors for neural radiance fields from sparse input views. In CVPR, pages 12892–12901, 2022. 1

    2022

  27. [35]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In CVPR, pages 10684– 10695, 2022. 3

    2022

  28. [36]

    Flipn- erf: Flipped reflection rays for few-shot novel view synthe- sis

    Seunghyeon Seo, Yeonjin Chang, and Nojun Kwak. Flipn- erf: Flipped reflection rays for few-shot novel view synthe- sis. In ICCV, pages 22883–22893, 2023. 2

    2023

  29. [37]

    Splatt3r: Zero-shot gaussian splat- ting from uncalibrated image pairs

    Brandon Smart, Chuanxia Zheng, Iro Laina, and Vic- tor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splat- ting from uncalibrated image pairs. In arXiv preprint arXiv:2408.13912, 2024. 3

    2024 arXiv

  30. [38]

    D ¨arf: Boosting radiance fields from sparse inputs with monocular depth adaptation

    Jiuhn Song, Seonghoon Park, Honggyu An, Seokju Cho, Min-Seop Kwak, Sungjin Cho, and Seungryong Kim. D ¨arf: Boosting radiance fields from sparse inputs with monocular depth adaptation. In NeurIPS, 2023. 2

    2023

  31. [39]

    Lgm: Large multi-view gaus- sian model for high-resolution 3d content creation

    Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaus- sian model for high-resolution 3d content creation. InECCV,

  32. [40]

    Mv-dust3r+: Single-stage scene reconstruction from sparse views in 2 seconds

    Zhenggang Tang, Yuchen Fan, Dilin Wang, Hongyu Xu, Rakesh Ranjan, Alexander Schwing, and Zhicheng Yan. Mv-dust3r+: Single-stage scene reconstruction from sparse views in 2 seconds. In CVPR, 2025. 3

    2025

  33. [41]

    Sparf: Neural radiance fields from sparse and noisy poses

    Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In CVPR, pages 4190–4200, 2023. 2

    2023

  34. [42]

    Scade: Nerfs from space carving with ambiguity-aware depth estimates

    Mikaela Angelina Uy, Ricardo Martin-Brualla, Leonidas Guibas, and Ke Li. Scade: Nerfs from space carving with ambiguity-aware depth estimates. In CVPR, 2023. 2

    2023

  35. [43]

    Ground- ing image matching in 3d with mast3r

    J ´erˆome Revaud Vincent Leroy, Yohann Cabon. Ground- ing image matching in 3d with mast3r. In arXiv preprint arXiv:2406.09756, 2024. 3, 4

    2024 arXiv

  36. [44]

    Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction

    Peng Wang, Hao Tan, Sai Bi, Yinghao Xu, Fujun Luan, Kalyan Sunkavalli, Wenping Wang, Zexiang Xu, and Kai Zhang. Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. In ICLR, 2024. 3, 6

    2024

  37. [45]

    Dust3r: Geometric 3d vi- sion made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vi- sion made easy. In CVPR, pages 20697–20709, 2024. 3, 4

    2024

  38. [46]

    Bovik, H.R

    Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600–612, 2004. 7

    2004

  39. [47]

    Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation

    Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In CVPR, pages 803–814, 2023. 6

    2023

  40. [48]

    Diffusionerf: Regularizing neural radiance fields with denoising diffusion models

    Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In CVPR, pages 4180–4189, 2023. 2

    2023

  41. [49]

    Sparp: Fast 3d object reconstruction and pose estimation from sparse views

    Chao Xu, Ang Li, Linghao Chen, Yulin Liu, Ruoxi Shi, Hao Su, and Minghua Liu. Sparp: Fast 3d object reconstruction and pose estimation from sparse views. InECCV, pages 143– 163, 2024. 3, 6

    2024

  42. [50]

    Freesplatter: Pose- free gaussian splatting for sparse-view 3d reconstruction

    Jiale Xu, Shenghua Gao, and Ying Shan. Freesplatter: Pose- free gaussian splatting for sparse-view 3d reconstruction. In arXiv preprint arXiv:2412.09573, 2024. 1, 2, 3, 4, 6

    2024 arXiv

  43. [51]

    Gaussianob- ject: High-quality 3d object reconstruction from four views with gaussian splatting

    Chen Yang, Sikuang Li, Jiemin Fang, Ruofan Liang, Lingxi Xie, Xiaopeng Zhang, Wei Shen, and Qi Tian. Gaussianob- ject: High-quality 3d object reconstruction from four views with gaussian splatting. ACM TOG, 43(6), 2024. 1, 2, 6

    2024

  44. [52]

    Freenerf: Im- proving few-shot neural rendering with free frequency regu- larization

    Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Im- proving few-shot neural rendering with free frequency regu- larization. In CVPR, pages 8254–8263, 2023. 2

    2023

  45. [53]

    Mvsnet: Depth inference for unstructured multi-view stereo

    Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In ECCV, pages 785–801, 2018. 2

    2018

  46. [54]

    No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images

    Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. In ICLR, 2025. 1, 3

    2025

  47. [55]

    pixelnerf: Neural radiance fields from one or few images

    Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, pages 4578–4587, 2021. 1, 2

    2021

  48. [56]

    Cor-gs: Sparse-view 3d gaussian splatting via co-regularization

    Jiawei Zhang, Jiahe Li, Xiaohan Yu, Lei Huang, Lin Gu, Jin Zheng, and Xiao Bai. Cor-gs: Sparse-view 3d gaussian splatting via co-regularization. In ECCV, pages 335–352,

  49. [57]

    Gs-lrm: Large recon- struction model for 3d gaussian splatting

    Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large recon- struction model for 3d gaussian splatting. In ECCV, pages 1–19, 2024. 1, 2, 3

    2024

  50. [58]

    Efros, Eli Shecht- man, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6, 7

    2018

  51. [59]

    Flare: Feed-forward geometry, appearance and camera estimation from uncalibrated sparse views

    Shangzhan Zhang, Jianyuan Wang, Yinghao Xu, Nan Xue, Christian Rupprecht, Xiaowei Zhou, Yujun Shen, and Gor- don Wetzstein. Flare: Feed-forward geometry, appearance and camera estimation from uncalibrated sparse views. In CVPR, pages 21936–21947, 2025. 1, 3

    2025

  52. [60]

    Sparse-view pose estima- tion and reconstruction via analysis by generative synthesis

    Qitao Zhao and Shubham Tulsiani. Sparse-view pose estima- tion and reconstruction via analysis by generative synthesis. In NeurIPS, 2024. 2 10

    2024

  53. [61]

    Bilateral refer- ence for high-resolution dichotomous image segmentation

    Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral refer- ence for high-resolution dichotomous image segmentation. CAAI Artificial Intelligence Research, 3:9150038, 2024. 1

    2024

  54. [62]

    Nexusgs: Sparse view synthesis with epipolar depth priors in 3d gaus- sian splatting

    Yulong Zheng, Zicheng Jiang, Shengfeng He, Yandu Sun, Junyu Dong, Huaidong Zhang, and Yong Du. Nexusgs: Sparse view synthesis with epipolar depth priors in 3d gaus- sian splatting. In CVPR, pages 26800–26809, 2025. 2

    2025

  55. [63]

    Is vanilla mlp in neural radiance field enough for few- shot view synthesis? In CVPR, pages 20288–20298, 2024

    Hanxin Zhu, Tianyu He, Xin Li, Bingchen Li, and Zhibo Chen. Is vanilla mlp in neural radiance field enough for few- shot view synthesis? In CVPR, pages 20288–20298, 2024. 2

    2024

  56. [64]

    Fsgs: Real-time few-shot view synthesis using gaussian splatting

    Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In ECCV, pages 145–163, 2024. 2

    2024

  57. [65]

    High- fidelity 3d textured shapes generation by sparse encoding and adversarial decoding

    Qi Zuo, Xiaodong Gu, Yuan Dong, Zhengyi Zhao, Weihao Yuan, Lingteng Qiu, Liefeng Bo, and Zilong Dong. High- fidelity 3d textured shapes generation by sparse encoding and adversarial decoding. In ECCV, pages 52–69, 2024. 2, 5, 6, 1 11 UFV-Splatter: Pose-Free Feed-Forward 3D Gau...

    2024

  58. [66]

    LoRA layers Table 4 presents a comparison of different configurations for inserting LoRA layers [17] on the GSO dataset

    Details of architecture 6.1. LoRA layers Table 4 presents a comparison of different configurations for inserting LoRA layers [17] on the GSO dataset. All LoRA layers are configured with a rank of 64. Since a large portion of the model consists of self-attention blocks, insert-...

  59. [67]

    Using these labeled Gaussians, we render both favorable target views and unfavorable context views

    Rendering details for unfavorable views To train the adapter, we label Gaussians for each sample from the G-Objaverse dataset [65], which contains only favorable-view images. Using these labeled Gaussians, we render both favorable target views and unfavorable context views. Fo...

  60. [68]

    3.4, we perform a linear alignment of the refined 3D Gaussians to enable rendering of target views during training

    Closed-form solution for alignment As described in Sec. 3.4, we perform a linear alignment of the refined 3D Gaussians to enable rendering of target views during training. This is formulated as a weighted least squares problem: min a,b NX i=1 X q∈Ω W c i (q) aeµi(q) + b − P c ...

  61. [69]

    Each example in- cludes the context views, a ground-truth target view, and novel-view results from FreeSplatter (w/ and w/o recenter- ing) [50], LucidFusion [14], and our method

    Additional results Figures 10 and 11 show additional qualitative comparisons on the GSO and OmniObject3D datasets. Each example in- cludes the context views, a ground-truth target view, and novel-view results from FreeSplatter (w/ and w/o recenter- ing) [50], LucidFusion [14],...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.