Predicting 3D representations for Dynamic Scenes

Beining Wang; Di Qi; Tong Yang; Wenqiang Zhang; Xiangyu Zhang

REVIEW 3 major objections 6 minor 1 cited by

Predicting 3D representations for Dynamic Scenes

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A single feed-forward model predicts the next moment's 3D radiance field from monocular video, with no scene-specific optimization, and generalizes to unseen dynamic scenes.

desk verdict A coherent next-3D prediction pipeline with honest ablations, but the central 3D claims outrun the image-space evidence. read the letter →

arxiv 2501.16617 v1 pith:755JSKKY submitted 2025-01-28 cs.CV

Di Qi , Tong Yang , Beining Wang , Xiangyu Zhang , Wenqiang Zhang This is my paper · ORCID

classification cs.CV

keywords dynamicradiancefieldpredictionego-centrictriplane4D-awaretransformermonocularvideoself-supervisedlearningnext-3Dnovelviewsynthesisgeneralizationtounseenscenes

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

This paper tries to establish that a single feed-forward network can take a short monocular video and predict an explicit 3D representation of the scene at the next time step, without optimizing anything on the target scene at inference. The representation is an ego-centric unbounded triplane centered on the camera, and a 4D-aware transformer aggregates the past frames' image features to update it. If true, this would give machines a form of 3D future prediction usable for robotics, driving, and augmented reality, and it would let large amounts of ordinary monocular video replace scarce multi-view 3D data for learning 4D priors. The reported results on dynamic-scene benchmarks and on unseen videos are the evidence offered for that claim.

What carries the argument

The load-bearing object is the ego-centric unbounded triplane: three axis-aligned feature planes anchored at the observer's camera, with unbounded coordinates mapped into a bounded box by the mip-NeRF 360 contraction. The 4D-aware transformer updates this triplane in three attention stages: temporal-aware view attention pulls epipolar features from the S source frames into 3D virtual point queries; axis attention and plane attention refine those point and plane features to repair the corrupted epipolar evidence that moving objects produce; camera intrinsics condition each block through adaptive layer normalization. A lightweight NeRF decoder turns the updated triplane into color and density, and a temporal-based 3D constraint renders two nearby target times from the same source sequence and applies photometric loss to enforce 3D consistency.

What would settle it

Render the predicted future triplane from viewpoints not seen in the source frames and compare dynamic-object depth against LiDAR or multi-view triangulation at t=S+1 in the NVIDIA scenes. If dynamic objects show systematic depth errors while image metrics look plausible, the claim that the model reconstructs true 3D motion is falsified.

Watch

Extended reading notes

Core claim

The paper claims that predicting a radiance field at a future time can be reduced to updating a learnable ego-centric triplane with cross-view and cross-time attention over monocular frames {I_t}, t=1..S, and then rendering the target view at t=S+1 through volume rendering. On NVIDIA Dynamic Scenes it reports full-image PSNR 22.43 against GNT's 20.33 and PGDVS†'s 19.23, with the largest margin in dynamic regions (18.64 vs 17.30 and 15.89); on DAVIS, a dataset outside its training distribution, it reports full-image PSNR 21.05 vs GNT's 19.10. The same training signal also produces, without depth supervision, predicted depth maps, and the image encoder's features support linear classification on ImageNet above a random baseline for categories related to training data. These observations are used to argue that the model learns generalized 4D priors and exhibits emergent geometry and semantic capabilities.

Load-bearing premise

The method's strong dynamic-region results depend on the untested premise that its attention modules can reconstruct correct 3D geometry for moving objects from noisy cross-view features alone, with no explicit motion estimation, optical flow, or object masks.

Editorial extensions

If this is right

If the central claim is right, dynamic 3D prediction no longer requires per-scene optimization or multi-view capture at inference: the same weights map past monocular frames to a future radiance field.
Rendering future novel views with this representation gives a concrete way to evaluate 3D future prediction using standard image metrics, since no ground-truth future 3D exists.
Training on abundant monocular video becomes a viable route to 4D priors; the reported generalization from kitchen and driving training data to unseen DAVIS scenes suggests the prior transfers across domains.
The emergent depth and linear-probing classification results imply the prediction objective itself can teach geometry and semantics, so self-supervised video pretraining of this kind may serve as a representation learner.
The method sets the 'next-3D prediction' task as distinct from future-frame prediction, giving later work a concrete setup: past frames in, future triplane out.

Reading between the lines

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

Beyond the paper, the rendering-metric evaluation leaves geometry unverified: a network could pass the image metrics while storing view-dependent textures rather than true 3D structure, so a direct depth check at t=S+1 would settle whether the representation is genuinely three-dimensional.
If attention-based repair of noisy dynamic epipolar features is the bottleneck, adding explicit motion cues such as optical flow or object masks should improve dynamic-region accuracy; the paper does not test this.
The paper lists its model's determinism as a limitation, pointing toward a natural next step of sampling multiple plausible futures from the same past frames with a generative model.
The paper notes that limited resources prevented training on larger egocentric datasets; if its reported generalization trend continues, scaling the training data should strengthen the emergent geometry and semantic signals.

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.

Desk Editor's Note

read the letter

Here is my read. The paper proposes a feed-forward model that takes S monocular frames with poses and predicts an ego-centric triplane for time S+1, trained self-supervised with photometric losses. The task formulation—predicting an explicit future 3D representation rather than a future frame—is genuinely new as far as I know, and the held-out evaluation on NVIDIA and DAVIS is a real step beyond per-scene optimization.

What the paper does well: the ego-centric unbounded triplane is a clean combination of contraction (mip-NeRF360) and triplanes (EG3D), and the 4D-aware transformer with temporal-aware view attention plus axis/plane attention is coherent. The ablations in Table 3 show each component helps, and the temporal-based 3D constraint gives a solid gain. Training only on monocular video with poses is the right scaling direction. The authors also list honest limitations in Appendix A.5.

Soft spots, in order of importance. First, the evaluation does not yet validate the central claim of 3D prediction. On DAVIS, the render is the same camera viewpoint at the next timestep, so strong PSNR/LPIPS can be achieved by a 2D video predictor that copies or lightly warps the last frame. On NVIDIA, the paper does not clearly show that the predicted triplane renders a held-out viewpoint at the target time, nor does it compare against actual video prediction baselines. The depth maps are qualitative only, and the Appendix's static/dynamic similarity analysis is a boxplot, not geometric validation. So 'predicts future 3D structure' is plausible but not yet demonstrated; the current evidence is consistent with a sophisticated image-synthesis shortcut.

Second, the baselines are weak and not fully matched. PGDVS† receives the target frame yet underperforms—worth explaining—but GNT is designed for static scenes. No error bars are reported, so some gaps, including the negative SSIM margin on DAVIS, may be noise. Third, the 'emergent semantic learning' claim overreaches: linear probing on ImageNet shows top categories tied to training data (odometer, school bus), which is not broad semantic understanding.

I agree with the stress-test note on the main gap. It is not a mathematical flaw, but it is load-bearing for the paper's headline. The fix is feasible: release code, evaluate view extrapolation and depth accuracy, add video prediction baselines, and report variance. I would send this to a serious referee, not desk-reject, because the task is new and the architecture is thoughtful. But I would ask for major revision, not acceptance as-is.

Referee Report

3 major / 6 minor

Summary. The paper proposes a feed-forward framework that, given a monocular video of S frames, predicts an ego-centric unbounded triplane representation of the scene at the next time step t=S+1. The triplane is updated by a 4D-aware transformer composed of temporal-aware view-attention, axis-attention, and plane-attention modules, and is trained self-supervised by rendering target views with a NeRF decoder and applying photometric, perceptual, and distortion losses plus a temporal-based 3D constraint. The method is evaluated on NVIDIA Dynamic Scenes and DAVIS, comparing against adapted versions of GNT and PGDVS, and shows higher PSNR/SSIM/LPIPS, especially in dynamic regions. The authors also report qualitative depth maps and ImageNet linear-probing results as evidence of emergent geometric and semantic capabilities.

Significance. If the claimed capability is genuine, the paper makes a useful contribution: it demonstrates that a generalizable future 3D representation can be learned from large-scale monocular video without scene-specific optimization, which is an important step toward 4D world modeling. The strengths include the self-supervised training recipe, the explicit triplane representation that supports rendering, the held-out evaluation on DAVIS, and ablations that support the roles of the proposed attention modules and losses. However, the current evidence does not yet establish that the predicted triplane encodes correct future 3D geometry; the evaluation is indirect and the evaluation protocol is under-specified. The emergent semantic results are preliminary but suggestive.

major comments (3)

[Section 4.1.1 (Evaluation setup)] The evaluation protocol does not state whether the target views on NVIDIA are rendered from camera viewpoints different from the source monocular sequence or from the same camera at the next time step. The text 'all next-time views can be treated as novel views' is compatible with both. If the target view is the same camera's next frame, the reported image-space metrics can be satisfied by a 2D video prediction model that warps or blends source appearance, and the results do not validate the predicted triplane as a 3D representation. If instead all 12 cameras at time S+1 are rendered from a single predicted triplane, the paper should state this explicitly and, ideally, report cross-view consistency or multi-view metrics, since this would directly counter the 2D-shortcut concern. As written, the central claim of '4D physical world modeling' is not supported by the reported experiments on DAVIS, which explicitly uses a single monocular view.
[Section 3.3, Eq. (4)] The temporal-based 3D constraint is not defined precisely. The text says the model predicts two 3D representations at time t=0 and t=S+1 and 'apply the photometric loss on these two rendered views,' but it is unclear what the supervision targets are: is the loss computed between each rendered view and its ground-truth frame, or between the two rendered views themselves? If the latter, the loss does not enforce agreement with any observed data and can be trivially minimized. Since Table 3 shows that removing this constraint causes a substantial drop (e.g., full-image PSNR from 28.56 to 25.78), the constraint is important, but the mechanism needs a precise mathematical statement.
[Section 3.2.2, Eq. (3)] The paper asserts that the axis-attention and plane-attention modules 'refine' the noisy epipolar features of dynamic content into correct 3D geometry, but it provides no direct evidence of geometric correctness. The only quantitative support is the similarity boxplot in Figure 7, which shows that attention scores differ between static and dynamic points; it does not show that the resulting triplane places dynamic content at the correct 3D location. To support the claim that the model predicts future 3D structure rather than a view-dependent image synthesis, the authors should provide geometric validation, e.g., depth maps of dynamic regions compared against ground-truth (multi-view depth or LiDAR) or rendering consistency across multiple target viewpoints at the same time. Without such evidence, the dynamic-area improvements in Tables 1 and 2 are also consistent with appearance-based image prediction.

minor comments (6)

[Section 3.3] The notation 't = 0' appears to be a typo; the temporal-based constraint likely refers to a source time such as t=S. Please correct and make the time indexing consistent with the problem definition {t}_{t=1}^S.
[Figure 7] Define the 'similarity' measure (e.g., cosine similarity between query and key features) and specify which attention layer or block the statistics are aggregated over.
[Tables 1-3] No error bars or significance tests are reported; since the linear-probing experiment in Appendix A.4 averages over three seeds, the same practice should be applied to the main quantitative claims.
[Section 4.1.1, DAVIS setup] 'evaluate novel view synthesis on a single view' is ambiguous. Please state whether this view is the same camera as the source sequence or a different one, and clarify how COLMAP poses are used.
[Table 1 and Figure 3] The symbol for the PGDVS baseline is inconsistent: 'PGDVS†' in Table 1 and text versus 'PGDVS+' in the Figure 3 caption; standardize the notation.
[Equation (2)] Please specify how the time embeddings for the query (time S+1) and keys (times 1..S) are constructed and injected; the text only mentions 'time embeddings' without a definition.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the model is trained self-supervised on held-out future frames and evaluated on unseen datasets; the unverified 3D intermediate is a validation gap, not a definitional reduction.

full rationale

The claimed derivation chain is not circular. The triplane representation is initialized as learnable parameters and updated by a transformer using source-frame features (Eqs. 2-3); the training objective (Eq. 4) is a photometric/perceptual reconstruction loss against ground-truth future frames, and the main evaluations (Tables 1-2) are on held-out NVIDIA Dynamic Scenes and DAVIS, not on the training set. No parameter is fitted to the test metric. The paper explicitly admits that the future 3D representation cannot be directly assessed and uses rendered-view metrics instead; this is an evaluation limitation, not a case where the prediction is equivalent to its inputs by construction. The static/dynamic attention analysis in Appendix A.2 uses ground-truth semantic masks as labels to verify emergent behavior, so it is an evaluation rather than a self-definition. There are no load-bearing self-citations: the cited prior work (mip-NeRF 360 contraction, plane-attention, upsampler) supplies components, not the central future-prediction claim. The 'emergent semantic learning' result is a linear-probing study whose highlighted categories correlate with training data, but the paper presents it as evidence of learned representation, not as the source of the main result. Overall, no circular step meeting the quoted-reduction bar was found.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

No extra physical entities are postulated. The free parameters are training hyperparameters, and the assumptions are mainly about representation sufficiency and the validity of the self-supervised training signal.

free parameters (8)

lambda_lpips = 0.1
Manual weight balancing perceptual loss; affects final LPIPS metric.
lambda_ssim = 0.1
Manual weight for structural similarity loss.
lambda_dist = 0.01
Manual weight for distortion regularization taken from mip-NeRF360.
learning_rate = 0.0001
Adam learning rate chosen for stable training.
transformer_feature_dim = 576
Channel width of image features, triplanes, and transformer.
num_transformer_blocks = 12
Depth of 4D-aware transformer stack.
triplane_spatial_resolution = 32x32 upsampled to 128x128
Triplane grid resolution before and after upsampling; chosen by hand.
camera_feature_dim = 16
Flattened camera intrinsic embedding dimension.

assumptions (5)

standard math Volume rendering with the NeRF decoder reconstructs target views from triplane features.
The photometric training loss assumes differentiable volume rendering is a valid image formation model.
standard math The mip-NeRF360 contraction maps unbounded 3D coordinates to bounded triplane coordinates.
Used to fit infinite scene space into a finite triplane, following Barron et al. 2022.
domain assumption Monocular videos with camera poses encode enough 3D information to learn 4D priors.
Section 1 justifies self-supervised training from monocular video on this basis.
domain assumption A single ego-centric triplane can represent arbitrary dynamic scenes in fixed size.
Section 3.1 assumes this representation transfers across scenes.
ad hoc to paper Two target views rendered from the same source sequence share dynamic scene overlap.
Stated explicitly in Section 3.3 as the assumption behind the temporal-based 3D constraint.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Predicting 3D representations for Dynamic Scenes." pith.science (2026). https://pith.science/paper/755JSKKY

@misc{pith2026250116617,
  author       = {Pith},
  title        = {Pith review of: Predicting 3D representations for Dynamic Scenes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/755JSKKY}},
  note         = {Machine review of arXiv:2501.16617}
}

read the original abstract

We present a novel framework for dynamic radiance field prediction given monocular video streams. Unlike previous methods that primarily focus on predicting future frames, our method goes a step further by generating explicit 3D representations of the dynamic scene. The framework builds on two core designs. First, we adopt an ego-centric unbounded triplane to explicitly represent the dynamic physical world. Second, we develop a 4D-aware transformer to aggregate features from monocular videos to update the triplane. Coupling these two designs enables us to train the proposed model with large-scale monocular videos in a self-supervised manner. Our model achieves top results in dynamic radiance field prediction on NVIDIA dynamic scenes, demonstrating its strong performance on 4D physical world modeling. Besides, our model shows a superior generalizability to unseen scenarios. Notably, we find that our approach emerges capabilities for geometry and semantic learning.

Figures

Figures reproduced from arXiv: 2501.16617 by the authors.

**Figure 1.** Overview of our method. The process starts with an image encoder that extracts 2D image features from the the video frames, serving as a prior. These features are then processed by a 4D-aware transformer, which updates a learnable triplane representation. Next, an upsampler refines and enlarges the triplane. During training, a NeRF decoder generates target views via volumetric rendering and photometric loss is appli… view at source ↗

**Figure 2.** Temporal-aware View-Attention and Axis-Attention modules in Transformer. (a) Temporal-aware View-Attention Module: At the target time t = S + 1, 3D virtual points are uniformly sampled within the triplane. For a given point xi,j,k, it is projected along the three axes onto the triplane features Txy, Tyz, and Txz to obtain the corresponding 3D query feature qi,j,k. Simultaneously, xi,j,k is mapped onto image feature … view at source ↗

**Figure 3.** Qualitative Comparison on the NVIDIA Dynamic Scenes Dataset. Our method significantly outperforms GNT and PGDVS† in both dynamic and static content. For dynamic objects (first two columns), our approach delivers precise motion and avoids motion blur compared with PGDVS† . For static contents (last two columns), our method shows a clear background, whereas PGDVS† and GNT result in blurred backgrounds due to limited d… view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: demonstrates the generalization capability of our framework on the DAVIS dataset. Although our model is trained only with kitchen and driving scenarios, it performs well on out-of-distribution scenes. In column 4, our model accurately predicts the motion of a blackswan…

**Figure 5.** Figure 5: Reconstructed depth maps on NVIDIA Dynamic Scenes. Red indicates closer distances, while blue denotes farther distances. 4.2 ABLATION STUDY In this section, we conduct ablation studies on the NVIDIA Dynamic Scenes dataset to better understand the contributions of diffe…

**Figure 6.** Figure 6: Comparisons on ImageNet linear classification: our model vs a random-initialized model. The highlighted categories (orange) are closely related to our training data. The categories in left and right parts are selected by top-1 classification accuracy of our model and a…

**Figure 7.** Figure 7: Boxplot summarizing the similarity measures of dynamic and static 3D points.The x-axis represents 6 views from the source frames, ordered by increasing temporal distance from the target view. A APPENDIX A.1 ARCHITECTURE DETAILS Image Encoder The image encoder consists …

**Figure 8.** Figure 8: Qualitative comparison of ablation studies on NVIDIA Dynamic Scenes dataset at 128 × 72 resolution. The metrics mean the PSNR on the testing dynamic and static contents in different ablation studies. A.4 COMPARISONS ON IMAGENET LINEAR CLASSIFICATION To evaluate the sem…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

FFAvatar: Feed-Forward 4D Head Avatar Reconstruction from Sparse Portrait Images
cs.CV 2026-06 unverdicted novelty 6.0 of 10

FFAvatar uses a Transformer-based 3D Gaussian model with alternating attention and sparse-to-dense learning to enable feed-forward, incremental reconstruction of animatable 4D head avatars from sparse portrait images.

Reference graph

Works this paper leans on

14 extracted references · 1 canonical work pages · cited by 1 Pith paper

[3]

Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang

11 Preprint. Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. arXiv preprint arXiv:2312.07504,

arXiv
[4]

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. arXiv preprint arXiv:2311.04400,

arXiv
[5]

Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models

Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763,

arXiv
[6]

Deformable neural radiance fields

Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Deformable neural radiance fields. arxiv. arXiv preprint arXiv:2011.12948,

arXiv 2011
[8]

The metrics mean the PSNR on the testing dynamic and static contents in different ablation studies

Static PSNR: 29.8 Dynamic PSNR: 24.3 Ours Static PSNR: 27.3 Dynamic PSNR: 20.7 Static PSNR: 28.6 Dynamic PSNR: 22.0 GT w/o Temporal-based 3D Constraint w/o Plane-Attentionw/o Self-Attention in Encoder Static PSNR: 29.0 Dynamic PSNR: 23.8 Static PSNR: 31.0 Dynamic PSNR: 25.1 w/o LPIPS Loss w/o SSIM Loss Static PSNR: 28.7 Dynamic PSNR: 23.1 Static PSNR: 29....

work page 2022
[11]

Emernerf: Emergent spatial-temporal scene decomposition via self-supervision

Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023a. Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasu...

arXiv
[12]

Dino: Detr with improved denoising anchor boxes for end-to-end object detection

Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605,

arXiv
[2004]

Multi-scale 3d gaussian splatting for anti-aliased rendering

Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. arXiv preprint arXiv:2311.17089,

arXiv

Show all 14 references

[2016]

Dreamfusion: Text-to-3d using 2d diffusion

Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988,

arXiv
[2018]

Xiaoming Zhao, Alex Colburn, Fangchang Ma, Miguel Angel Bautista, Joshua M

13 Preprint. Xiaoming Zhao, Alex Colburn, Fangchang Ma, Miguel Angel Bautista, Joshua M. Susskind, and Alexander G. Schwing. Pseudo-Generalized Dynamic View Synthesis from a Video. In ICLR, 2024a. Xiaoming Zhao, R Alex Colburn, Fangchang Ma, Miguel Ángel Bautista, Joshua M Sus...

arXiv
[2020]

Fourier plenoctrees for dynamic radiance field rendering in real-time

Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp....

arXiv
[2021]

Splatter image: Ultra-fast single-view 3d reconstruction

Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. arXiv preprint arXiv:2312.13150,

arXiv
[2022]

Zoedepth: Zero-shot transfer by combining relative and metric depth

Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288,

arXiv
[2023]

Large-vocabulary 3d diffusion model with transformer

Ziang Cao, Fangzhou Hong, Tong Wu, Liang Pan, and Ziwei Liu. Large-vocabulary 3d diffusion model with transformer. arXiv preprint arXiv:2309.07920,

arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [3]

Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang

11 Preprint. Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. arXiv preprint arXiv:2312.07504,

arXiv

[2] [4]

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. arXiv preprint arXiv:2311.04400,

arXiv

[3] [5]

Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models

Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763,

arXiv

[4] [6]

Deformable neural radiance fields

Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Deformable neural radiance fields. arxiv. arXiv preprint arXiv:2011.12948,

arXiv 2011

[5] [8]

The metrics mean the PSNR on the testing dynamic and static contents in different ablation studies

Static PSNR: 29.8 Dynamic PSNR: 24.3 Ours Static PSNR: 27.3 Dynamic PSNR: 20.7 Static PSNR: 28.6 Dynamic PSNR: 22.0 GT w/o Temporal-based 3D Constraint w/o Plane-Attentionw/o Self-Attention in Encoder Static PSNR: 29.0 Dynamic PSNR: 23.8 Static PSNR: 31.0 Dynamic PSNR: 25.1 w/o LPIPS Loss w/o SSIM Loss Static PSNR: 28.7 Dynamic PSNR: 23.1 Static PSNR: 29....

work page 2022

[6] [11]

Emernerf: Emergent spatial-temporal scene decomposition via self-supervision

Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023a. Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasu...

arXiv

[7] [12]

Dino: Detr with improved denoising anchor boxes for end-to-end object detection

Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605,

arXiv

[8] [2004]

Multi-scale 3d gaussian splatting for anti-aliased rendering

Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. arXiv preprint arXiv:2311.17089,

arXiv