Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

WonderFree: Enhancing Novel View Quality and Cross-View Consistency for 3D Scene Exploration

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read WonderFree restores novel-view videos so users can explore 3D worlds from any angle.

desk verdict A plausible systems paper whose headline claim about cross-view consistency is not backed by the metrics it reports; the evaluation needs geometric checks before the central loop can be trusted. read the letter →

arxiv 2506.20590 v1 pith:7665YUL4 submitted 2025-06-25 cs.CV

classification cs.CV
keywords 3Dworldgenerationnovelviewsynthesisvideorestorationmulti-viewconsistencyGaussianSplattinginteractivesceneexploration
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

The paper puts forward WonderFree, an interactive system for turning a single image into a 3D world that a user can walk through from any angle, including forward motion into previously unseen regions. The central claim is that two linked failures—floater-filled renderings on novel views and disagreement between neighboring viewpoints—can both be fixed by a restore-and-refine loop: render novel-view videos from a coarse 3D world, clean them with a learned video-restoration model, and feed the cleaned videos back into the world representation. To train that restorer, the authors build a data pipeline that pairs degraded renderings of under-trained Gaussian splatting models with clean ground truth, across real and synthetic indoor, urban, and nature scenes. They report that the resulting system raises CLIP Score from 32.28 to 35.00 and wins 77.20% of user comparisons against the previous state of the art, WonderWorld. The paper matters because unrestricted exploration is the main missing capability in interactive 3D scene generation.

What carries the argument

The load-bearing mechanism is the render-restore-refine loop. A coarse 3D Gaussian Splatting world is rendered from novel trajectories; the degraded video goes into WorldRestorer, a conditional video-denoising diffusion model initialized from a pretrained video generator, which reconstructs clean frames; ConsistView pre-aligns multiple viewpoints (samples at angular offsets $\theta_k = (k - n)\Delta\theta$ around the forward direction) into one wide image so the restorer treats them jointly, preserving spatiotemporal coherence; and the restored videos are used to refine the splatting world. The loop feeds on WorldScopeDataset, a hybrid real-and-synthetic corpus in which degraded videos are produced by rendering under-trained Gaussian splatting checkpoints along camera trajectories and paired with clean ground-truth frames, with artificial masks simulating occlusions.

What would settle it

Render novel-view videos from a fully trained interactive 3D world along forward trajectories the world was never shown, restore them with WorldRestorer, and compare the restored frames against independently captured ground-truth images of the same real scene (or against the ground-truth render of a synthetic scene built with the game-engine pipeline). If the restored frames contain structure or color that contradicts the independent capture at a high rate, the under-trained-checkpoint training assumption fails and the refinement loop would be expected to worsen the world rather than repair it.

Watch

Extended reading notes

Core claim

On its own terms, WonderFree's discovery is that the limited explorability of generated 3D worlds is not a single rendering problem but a coupled quality-and-consistency problem that can be solved jointly by a data-driven video restorer embedded in an iterative world-refinement loop. The system first builds a coarse 3D Gaussian Splatting world, then renders videos along novel camera trajectories—including forward, diagonal, and lateral moves—that are corrupted by floaters and ghosting. WorldRestorer, a conditional video denoising network fine-tuned from a pretrained video generation model, removes those artifacts, and ConsistView makes the restoration multi-view-aware by horizontally concatenating viewpoints captured at fixed angular offsets, so lighting, geometry, and content stay coherent across simultaneously rendered perspectives. The cleaned videos then become pseudo-ground-truth supervision for a world-refinement step, and the loop repeats. The paper argues that this loop, trained on the WorldScopeDataset of paired degraded/clean multi-view videos, is what lets WonderFree beat WonderWorld on every reported metric and win the user study.

Load-bearing premise

WorldRestorer is trained on artifacts from deliberately under-trained Gaussian splatting models, and everything depends on those artifacts resembling what actually appears in novel-view renderings of interactively generated worlds; if the two artifact distributions diverge, the restorer will hallucinate instead of repair, and re-feeding its outputs into the world refinement will bake new errors into the scene.

Editorial extensions

If this is right

  • Users can be expected to move through generated worlds with forward, diagonal, and lateral camera paths—not just pan around a fixed point—and still see clean renderings.
  • Neighboring viewpoints agree on geometry, lighting, and texture because restoration is done jointly across views, not per-image.
  • Each pass of the restore-and-refine loop improves the coarse world, so longer exploration yields progressively better scenes.
  • The hybrid data pipeline (real scans and game-engine synthetic scenes) gives the restorer coverage across indoor, urban, and nature styles, making the approach work beyond a single domain.

Reading between the lines

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

  • If the loop works as claimed, the same render-restore-refine design should transfer to other 3D representations such as meshes or NeRFs, where the coarse stage also suffers from floaters in unseen regions.
  • ConsistView's trick of stitching angularly offset views into a single canvas is a general recipe for multi-camera video restoration and could apply to VR passthrough, telepresence, or multi-camera video enhancement outside 3D generation.
  • A stress test the paper does not report is whether the user preference persists when the input image or scene style lies far outside the training distribution of WorldScopeDataset; that is the main empirical risk to the exploration claim.
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

4 major / 5 minor

Summary. The paper presents WonderFree, an iterative framework for improving novel-view quality and cross-view consistency in single-image 3D scene generation. A coarse 3D Gaussian Splatting world is rendered along novel trajectories; a video-restoration model (WorldRestorer) enhanced with a multi-view joint-restoration mechanism (ConsistView) cleans the rendered videos; the restored videos are then used as pseudo-ground-truth to refine the 3D world. Training data are generated from under-trained 3DGS renderings and synthetic Unreal Engine sequences (WorldScopeDataset). Experiments report improvements over LucidDreamer, Text2Room, DreamScene360, WonderJourney, WonderWorld, and WonderTurbo on CLIP Score, CLIP Consistency, CIQA, Q-Align, CLIP Aesthetic, and a user study.

Significance. If the method performs as claimed, it would be a useful step toward interactive exploration of generated 3D scenes, and the automated construction of a restoration dataset is a practically valuable contribution. The main strengths are the clear separation of novel-view quality and cross-view consistency, the data-generation pipeline, and the promise of public code/model/data. However, the current evidence does not yet substantiate the cross-view consistency claim because no geometric evaluation is used, and the evaluation scale and statistical rigor are insufficient for the strength of the claims.

major comments (4)
  1. [Sec 3.3, Eq. (2); Sec 4.2, Table 2] The central claim of cross-view consistency is not tested by any geometric metric. ConsistView concatenates K views into a wide image and restores them jointly, which can encourage appearance agreement, but neither this mechanism nor the training objective in Eq. (2) enforces epipolar, depth, or reprojection consistency. The 'CLIP Consistency' metric of Sec 4.2 and Appendix B compares CLIP embedding similarities to a central reference view, which is a semantic/appearance measure, not a 3D-consistency measure. Because the restored videos are used as pseudo-ground-truth for world refinement (Sec 3.1), unvalidated geometric errors can be baked into the 3D world. I recommend adding a geometric evaluation (e.g., depth consistency or multi-view reprojection error on the rendered novel views) and, if possible, a geometric consistency loss in training or a filtering criterion before refinement.
  2. [Sec 3.4, Eq. (5)] The training data for WorldRestorer are rendered videos from under-trained 3DGS models replayed along the original trajectory C. At inference, the model is applied to renderings from novel trajectories of interactively generated 3D worlds. The paper does not provide evidence that the artifact distribution in these two settings matches; if there is a domain gap, the restored pseudo-GT frames used in the refinement loop (Sec 3.2) could hallucinate content rather than repair it. A concrete test would be to run WorldRestorer on corrupted renderings from a held-out set of generated worlds and measure whether restoration reduces floaters while preserving geometry, e.g., by comparing the restored frames to ground-truth renderings of the same scene where available.
  3. [Sec 4.2, Table 2; Implementation Details] All quantitative conclusions rest on 32 scenes with no reported error bars, confidence intervals, or significance tests. Gains such as CLIP Score 35.00 vs 32.28 and Q-Align 3.912 vs 3.437 may be within run-to-run variance; the paper should report per-scene standard deviations and a paired significance test across scenes. The user study (Table 3) also needs the number of participants, the number of comparisons per participant, and the protocol to interpret the win rates.
  4. [Sec 4.2, User Study paragraph] The text states that WonderFree achieves 'comparable performance' to WonderWorld [1] and WonderTurbo [8], while Table 3 reports win rates of 77.20% and 78.40% for WonderFree. These statements are contradictory; a win rate of about 77% is a large preference margin, not 'comparable'. This should be corrected, and the protocol clarified.
minor comments (5)
  1. [Table 3] The method identifiers in parentheses appear swapped: WonderWorld is reference [1] and WonderTurbo is reference [8], but the table lists 'vs. WonderWorld [24]' and 'vs. WonderTurbo [1]'.
  2. [Abstract and Sec 1] The claim of 'arbitrary angles and directions' is stronger than the five trajectory types listed in Appendix B; I suggest softening to 'varied exploration directions' unless additional trajectory coverage is demonstrated.
  3. [Fig. 2 caption] The word 'Trajrctory' is misspelled, and it appears twice in the caption.
  4. [Sec 3.2 and Sec 3.4] Notation: T is used for the number of 3DGS checkpoints in Eq. (5) and as the control condition in Eq. (2); please disambiguate these uses.
  5. [Table 1] Please clarify whether the 23.4M images count includes both degraded and clean frames, and how Dreal/Ssynth relate to the reported scene count.

Circularity Check

1 steps flagged · score 6.0 of 10

Novel-view evaluation is performed on the same trajectories used to supervise world refinement, so the reported gains reduce to in-sample fitting of restored pseudo-ground truth.

  1. fitted input called prediction [Sec. 3.1 (Refinement loop) + Sec. 4.1 (Implementation Details) + App. B (Camera Trajectories)]
    "In the subsequent world refinement stage, the restored videos serve as supervisory signals to iteratively improve both the geometric fidelity and multi-view consistency of the generated 3D world. ... To guarantee consistency across comparisons, we utilize the same camera configuration throughout scene generation and evaluation. ... During testing, we select views at regular intervals along each trajectory as novel views, and the final evaluation metrics are computed by averaging results across all paths."

    The refinement loop renders videos from 'new trajectories', restores them with WorldRestorer, and uses those restored videos as supervision to refine the 3D world. The quantitative evaluation then selects views 'at regular intervals along each trajectory' and averages metrics over 'all paths', with the same camera configuration used in scene generation and evaluation. Thus the 'novel views' in Tables 2 and 4 are the very views whose restored versions were used as target signals to fit the world during WonderFree's per-scene refinement. The reported CLIP/Q-Align improvements over WonderWorld are in-sample measures of how well the refined world reproduces the restoration model's outputs on those trajectories, not of generalization to unseen viewpoints.

full rationale

The main circularity is in the evaluation protocol, not in the training equations. WorldRestorer is trained on an externally constructed dataset (WorldScopeDataset) with a standard conditional diffusion loss, so that part is self-contained. The per-scene refinement loop, however, renders novel-trajectory videos, restores them, and then treats the restored videos as supervision for refining the 3D world. The paper's quantitative evaluation uses the same camera configuration and selects evaluation views along trajectories without stating that these are held out from the refinement trajectories. Consequently, Tables 2 and 4 for WonderFree measure how well the refined world reproduces the pseudo-ground-truth restored frames on the very trajectories used for refinement, making the reported 'novel view' gains partially forced. The user study is an independent preference signal and prevents a higher score, but it does not rescue the quantitative claim of cross-view consistency from this in-sample evaluation. Self-citations such as WonderTurbo appear as baselines but are not load-bearing for the derivation, and no uniqueness theorem or ansatz is imported from the authors' prior work.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central pipeline rests on several domain assumptions rather than formal derivations. Most load-bearing are the artifact-distribution match between under-trained 3DGS renders and novel-view renders, the usefulness of SVD as a restoration prior, and the claim that side-by-side joint denoising yields multi-view consistency. The free parameters are mostly experimental choices (view count, angular step, checkpoint epochs, masks) whose values are not fully reported.

free parameters (4)
  • Angular step Delta theta and view count K in ConsistView = not reported
    Eqs. (3)-(4) define K = 2n+1 views separated by Delta theta; these determine overlap and consistency, but no values are given.
  • 3DGS under-training checkpoint epochs = around 500-2500 epochs (Fig 7)
    Degraded videos are sampled from randomly saved checkpoints; artifact types are defined by these epochs.
  • Spatial mask M = not specified
    M is applied to degraded frames to simulate occlusions (Sec 3.2); size, placement, and ratio are unspecified.
  • Number of restoration checkpoints T and trajectory segment count N = not reported
    Dreal uses T saved checkpoints and N equal-length segments (Sec 3.4); dataset scale depends on these numbers.
assumptions (5)
  • domain assumption Rendering under-trained 3DGS checkpoints of real videos approximates artifact distributions in novel views of generated worlds.
    Foundation of the WorldScopeDataset training pairs (Sec 3.4).
  • domain assumption Stable Video Diffusion gives a strong video prior for restoration fine-tuning.
    WorldRestorer is initialized from SVD (Sec 3.2).
  • domain assumption Side-by-side concatenation and joint denoising produce multi-view consistency.
    ConsistView (Sec 3.3) has no geometric constraint.
  • domain assumption CLIP-based metrics can measure novel-view quality and cross-view consistency.
    All quantitative conclusions come from these metrics (Sec 4.1, Appendix B).
  • domain assumption Restored pseudo-GT videos can serve as reliable supervision for world refinement.
    Refinement loop (Sec 3.1, 3.2) treats restoration outputs as ground truth.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WonderFree: Enhancing Novel View Quality and Cross-View Consistency for 3D Scene Exploration." pith.science (2026). https://pith.science/paper/7665YUL4

@misc{pith2026250620590,
  author       = {Pith},
  title        = {Pith review of: WonderFree: Enhancing Novel View Quality and Cross-View Consistency for 3D Scene Exploration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7665YUL4}},
  note         = {Machine review of arXiv:2506.20590}
}
read the original abstract

Interactive 3D scene generation from a single image has gained significant attention due to its potential to create immersive virtual worlds. However, a key challenge in current 3D generation methods is the limited explorability, which cannot render high-quality images during larger maneuvers beyond the original viewpoint, particularly when attempting to move forward into unseen areas. To address this challenge, we propose WonderFree, the first model that enables users to interactively generate 3D worlds with the freedom to explore from arbitrary angles and directions. Specifically, we decouple this challenge into two key subproblems: novel view quality, which addresses visual artifacts and floating issues in novel views, and cross-view consistency, which ensures spatial consistency across different viewpoints. To enhance rendering quality in novel views, we introduce WorldRestorer, a data-driven video restoration model designed to eliminate floaters and artifacts. In addition, a data collection pipeline is presented to automatically gather training data for WorldRestorer, ensuring it can handle scenes with varying styles needed for 3D scene generation. Furthermore, to improve cross-view consistency, we propose ConsistView, a multi-view joint restoration mechanism that simultaneously restores multiple perspectives while maintaining spatiotemporal coherence. Experimental results demonstrate that WonderFree not only enhances rendering quality across diverse viewpoints but also significantly improves global coherence and consistency. These improvements are confirmed by CLIP-based metrics and a user study showing a 77.20% preference for WonderFree over WonderWorld enabling a seamless and immersive 3D exploration experience. The code, model, and data will be publicly available.

Figures

Figures reproduced from arXiv: 2506.20590 by the authors.

Figure 1
Figure 1. We compare our method with WonderWorld [ [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the WonderFree pipeline. WonderFree first builds a 3D world, then renders [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Examples of degraded video frames and their restored counterparts. WorldRestorer with [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: The process of collecting data in Unreal Engine. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparisons between WonderFree and WonderWorld [ [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Visualization comparison of WorldRestorer with and without ConsistView. The first, [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: The video restoration dataset for WorldRestorer training, where images rendered by 3DGS [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Qualitative examples. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Qualitative examples. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: Qualitative examples. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: Qualitative examples. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 12
Figure 12. Figure 12: Qualitative examples. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_12.png]
Figure 13
Figure 13. Figure 13: Qualitative examples. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_13.png]
Figure 14
Figure 14. Figure 14: Qualitative examples. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_14.png]

Discussion (0). Sign in 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

  1. GigaWorld-1: A Roadmap to Build World Models for Robot Policy Evaluation

    cs.RO 2026-07 conditional novelty 6.0 of 10

    Long-horizon action-faithful consistency, not short-term visual realism, dominates world-model reliability for robot policy evaluation; GigaWorld-1 implements that roadmap and gains 14.9% on evaluator-alignment metrics.

Reference graph

Works this paper leans on

61 extracted references · 25 canonical work pages · cited by 1 Pith paper

  1. [1]

    Wonder- world: Interactive 3d scene generation from a single image

    Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T Freeman, and Jiajun Wu. Wonder- world: Interactive 3d scene generation from a single image. arXiv preprint arXiv:2406.09394, 2024

  2. [8]

    Wonderturbo: Generating interactive 3d world in 0.72 seconds

    Chaojun Ni, Xiaofeng Wang, Zheng Zhu, Weijie Wang, Haoyun Li, Guosheng Zhao, Jie Li, Wenkang Qin, Guan Huang, and Wenjun Mei. Wonderturbo: Generating interactive 3d world in 0.72 seconds. arXiv preprint arXiv:2504.02261, 2025

  3. [2]

    Text2room: Extracting textured 3d meshes from 2d text-to-image models

    Lukas Höllein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909–7920, 2023

  4. [3]

    Lu- ciddreamer: Domain-free generation of 3d gaussian splatting scenes

    Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Lu- ciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023

  5. [4]

    Genex: Generating an explorable world

    Taiming Lu, Tianmin Shu, Junfei Xiao, Luoxin Ye, Jiahao Wang, Cheng Peng, Chen Wei, Daniel Khashabi, Rama Chellappa, Alan Yuille, et al. Genex: Generating an explorable world. arXiv preprint arXiv:2412.09624, 2024

  6. [5]

    Pano2room: Novel view synthesis from a single indoor panorama

    Guo Pu, Yiming Zhao, and Zhouhui Lian. Pano2room: Novel view synthesis from a single indoor panorama. In SIGGRAPH Asia 2024 Conference Papers, pages 1–11, 2024

  7. [6]

    Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting

    Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Tejas Bharadwaj, Suya You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In European Conference on Computer Vision, pages 324–342. Springer, 2024

  8. [7]

    Diffpano: Scalable and consistent text to panorama generation with spherical epipolar-aware diffusion

    Weicai Ye, Chenhao Ji, Zheng Chen, Junyao Gao, Xiaoshui Huang, Song-Hai Zhang, Wanli Ouyang, Tong He, Cairong Zhao, and Guofeng Zhang. Diffpano: Scalable and consistent text to panorama generation with spherical epipolar-aware diffusion. Advances in Neural Information Processing Systems, 37:1304–1332, 2025

Show all 61 references
  1. [9]

    Diffusion360: Seamless 360 de- gree panoramic image generation based on diffusion models

    Mengyang Feng, Jinlin Liu, Miaomiao Cui, and Xuansong Xie. Diffusion360: Seamless 360 de- gree panoramic image generation based on diffusion models. arXiv preprint arXiv:2311.13141, 2023

  2. [10]

    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 European Conference on Computer Vision, pages 370–386. Springer, 2024

  3. [11]

    Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis

    Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024

  4. [12]

    Layer-structured 3d scene inference via view synthesis

    Shubham Tulsiani, Richard Tucker, and Noah Snavely. Layer-structured 3d scene inference via view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), pages 302–317, 2018

  5. [13]

    3d photography using context- aware layered depth inpainting

    Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context- aware layered depth inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8028–8038, 2020

  6. [14]

    Dust3r: Geometric 3d vision made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697–20709, 2024. 10

  7. [15]

    Single-view view synthesis with multiplane images

    Richard Tucker and Noah Snavely. Single-view view synthesis with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 551–560, 2020

  8. [16]

    Synsin: End-to-end view synthesis from a single image

    Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7467–7477, 2020

  9. [17]

    Worldscore: A unified evaluation benchmark for world generation

    Haoyi Duan, Hong-Xing Yu, Sirui Chen, Li Fei-Fei, and Jiajun Wu. Worldscore: A unified evaluation benchmark for world generation. arXiv preprint arXiv:2504.00983, 2025

  10. [18]

    Sculpt3d: Multi-view consistent text-to-3d generation with sparse 3d prior

    Cheng Chen, Xiaofeng Yang, Fan Yang, Chengzeng Feng, Zhoujie Fu, Chuan-Sheng Foo, Guosheng Lin, and Fayao Liu. Sculpt3d: Multi-view consistent text-to-3d generation with sparse 3d prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, page...

  11. [19]

    Gaussctrl: Multi-view consistent text-driven 3d gaussian splatting editing

    Jing Wu, Jia-Wang Bian, Xinghui Li, Guangrun Wang, Ian Reid, Philip Torr, and Victor Adrian Prisacariu. Gaussctrl: Multi-view consistent text-driven 3d gaussian splatting editing. In European Conference on Computer Vision, pages 55–71. Springer, 2024

  12. [20]

    Sv4d: Dy- namic 3d content generation with multi-frame and multi-view consistency

    Yiming Xie, Chun-Han Yao, Vikram V oleti, Huaizu Jiang, and Varun Jampani. Sv4d: Dy- namic 3d content generation with multi-frame and multi-view consistency. arXiv preprint arXiv:2407.17470, 2024

  13. [21]

    Viewfusion: Towards multi-view consistency via interpolated denoising

    Xianghui Yang, Yan Zuo, Sameera Ramasinghe, Loris Bazzani, Gil Avraham, and Anton van den Hengel. Viewfusion: Towards multi-view consistency via interpolated denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9870–9880, 2024

  14. [22]

    Multi-view consistency as supervisory signal for learning shape and pose prediction

    Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Multi-view consistency as supervisory signal for learning shape and pose prediction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2897–2905, 2018

  15. [23]

    Drivedreamer-2: Llm-enhanced world models for diverse driving video generation

    Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Xinze Chen, Guan Huang, Xiaoyi Bao, and Xin- gang Wang. Drivedreamer-2: Llm-enhanced world models for diverse driving video generation. arXiv preprint arXiv:2403.06845, 2024

  16. [24]

    Wonderjourney: Going from anywhere to everywhere

    Hong-Xing Yu, Haoyi Duan, Junhwa Hur, Kyle Sargent, Michael Rubinstein, William T Freeman, Forrester Cole, Deqing Sun, Noah Snavely, Jiajun Wu, et al. Wonderjourney: Going from anywhere to everywhere. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Rec...

  17. [25]

    Stable video diffusion: Scaling latent video diffusion models to large datasets

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Do- minik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023

  18. [26]

    Plataniotis, Sergey Tulyakov, and Jian Ren

    Hanwen Liang, Junli Cao, Vidit Goel, Guocheng Qian, Sergei Korolev, Demetri Terzopoulos, Konstantinos N. Plataniotis, Sergey Tulyakov, and Jian Ren. Wonderland: Navigating 3d scenes from a single image, 2024

  19. [27]

    Recondreamer: Crafting world models for driving scene reconstruction via online restoration

    Chaojun Ni, Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Wenkang Qin, Guan Huang, Chen Liu, Yuyin Chen, Yida Wang, Xueyang Zhang, et al. Recondreamer: Crafting world models for driving scene reconstruction via online restoration. arXiv preprint arXiv:2411.19548, 2024

  20. [28]

    Recondreamer++: Harmonizing generative and reconstructive models for driving scene representation

    Guosheng Zhao, Xiaofeng Wang, Chaojun Ni, Zheng Zhu, Wenkang Qin, Guan Huang, and Xingang Wang. Recondreamer++: Harmonizing generative and reconstructive models for driving scene representation. arXiv preprint arXiv:2503.18438, 2025

  21. [29]

    Human- dreamer: Generating controllable human-motion videos via decoupled generation

    Boyuan Wang, Xiaofeng Wang, Chaojun Ni, Guosheng Zhao, Zhiqin Yang, Zheng Zhu, Muyang Zhang, Yukun Zhou, Xinze Chen, Guan Huang, Lihong Liu, and Xingang Wang. Human- dreamer: Generating controllable human-motion videos via decoupled generation. arXiv preprint arXiv:2503.24026,...

  22. [30]

    Humandreamer-x: Photorealistic single-image human avatars reconstruction via gaussian restoration

    Boyuan Wang, Runqi Ouyang, Xiaofeng Wang, Zheng Zhu, Guosheng Zhao, Chaojun Ni, Guan Huang, Lihong Liu, and Xingang Wang. Humandreamer-x: Photorealistic single-image human avatars reconstruction via gaussian restoration. arXiv preprint arXiv:2504.03536, 2025

  23. [31]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

  24. [32]

    Cogvideo: Large-scale pretraining for text-to-video generation via transformers

    Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022

  25. [33]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139–1, 2023

  26. [34]

    Drivedreamer4d: World models are effective data machines for 4d driving scene representation

    Guosheng Zhao, Chaojun Ni, Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Boyuan Wang, Youyi Zhang, Wenjun Mei, and Xingang Wang. Drivedreamer4d: World models are effective data machines for 4d driving scene representation. arXiv preprint arXiv:2410.13571, 2024

  27. [35]

    Sgd: Street view synthesis with gaussian splatting and diffusion prior

    Zhongrui Yu, Haoran Wang, Jinze Yang, Hanzhang Wang, Jiale Cao, Zhong Ji, and Mingming Sun. Sgd: Street view synthesis with gaussian splatting and diffusion prior. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3812–3822. IEEE, 2025

  28. [36]

    3dgs-enhancer: Enhancing unbounded 3d gaussian splatting with view-consistent 2d diffusion priors

    Xi Liu, Chaoyi Zhou, and Siyu Huang. 3dgs-enhancer: Enhancing unbounded 3d gaussian splatting with view-consistent 2d diffusion priors. Advances in Neural Information Processing Systems, 37:133305–133327, 2024

  29. [37]

    From an image to a scene: Learning to imagine the world from a million 360° videos

    Matthew Wallingford, Anand Bhattad, Aditya Kusupati, Vivek Ramanujan, Matt Deitke, Anirud- dha Kembhavi, Roozbeh Mottaghi, Wei-Chiu Ma, and Ali Farhadi. From an image to a scene: Learning to imagine the world from a million 360° videos. Advances in Neural Information Processin...

  30. [38]

    Drive- dreamer: Towards real-world-driven world models for autonomous driving

    Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drive- dreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023

  31. [39]

    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. Advances in Neural Information Processing Systems, 37:107064–107086, 2025

  32. [40]

    Evagaussians: Event stream assisted gaussian splatting from blurry images

    Wangbo Yu, Chaoran Feng, Jiye Tang, Jiashu Yang, Zhenyu Tang, Xu Jia, Yuchao Yang, Li Yuan, and Yonghong Tian. Evagaussians: Event stream assisted gaussian splatting from blurry images. arXiv preprint arXiv:2405.20224, 2024

  33. [42]

    Indoor segmentation and support inference from rgbd images

    Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision–ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746–760. Spr...

  34. [43]

    Scenenn: A scene meshes dataset with annotations

    Binh-Son Hua, Quang-Hieu Pham, Duc Thanh Nguyen, Minh-Khoi Tran, Lap-Fai Yu, and Sai-Kit Yeung. Scenenn: A scene meshes dataset with annotations. In 2016 fourth international conference on 3D vision (3DV), pages 92–101. Ieee, 2016

  35. [44]

    Scannet: Richly-annotated 3d reconstructions of indoor scenes

    Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017. 12

  36. [45]

    The replica dataset: A digital replica of indoor spaces

    Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019

  37. [46]

    nuscenes: A multimodal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020

  38. [47]

    Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding

    Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF international conference on ...

  39. [48]

    3d-front: 3d furnished rooms with layouts and semantics

    Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10933– 10...

  40. [49]

    Aligning latent and image spaces to connect the unconnectable

    Ivan Skorokhodov, Grigorii Sotnikov, and Mohamed Elhoseiny. Aligning latent and image spaces to connect the unconnectable. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14144–14153, 2021

  41. [50]

    Infinite nature: Perpetual view generation of natural scenes from a single image

    Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458– 14467, 2021

  42. [51]

    Scannet++: A high- fidelity dataset of 3d indoor scenes

    Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high- fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12–22, 2023

  43. [52]

    3d scene generation: A survey

    Beichen Wen, Haozhe Xie, Zhaoxi Chen, Fangzhou Hong, and Ziwei Liu. 3d scene generation: A survey. arXiv preprint arXiv:2505.05474, 2025

  44. [53]

    Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision

    Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, 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, pa...

  45. [54]

    Scannet: Richly-annotated 3d reconstructions of indoor scenes

    Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017

  46. [55]

    A benchmark for the evaluation of rgb-d slam systems

    Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of rgb-d slam systems. In2012 IEEE/RSJ international conference on intelligent robots and systems, pages 573–580. IEEE, 2012

  47. [56]

    Are we ready for autonomous driving? the kitti vision benchmark suite

    Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354–3361. IEEE, 2012

  48. [57]

    Unidentified video objects: A benchmark for dense, open-world segmentation

    Weiyao Wang, Matt Feiszli, Heng Wang, and Du Tran. Unidentified video objects: A benchmark for dense, open-world segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10776–10785, 2021

  49. [58]

    Structure-from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceed- ings of the IEEE conference on computer vision and pattern recognition , pages 4104–4113, 2016

  50. [59]

    Unreal engine

    Unreal Engine. Unreal engine. Retrieved from Unreal Engine: https://www. unrealengine. com/en-US/what-is-unreal-engine-4, 2018. 13

  51. [60]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021

  52. [61]

    Exploring clip for assessing the look and feel of images

    Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 2555–2563, 2023

  53. [62]

    Q-align: Teaching lmms for visual scoring via discrete text-defined levels

    Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching lmms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 14 A The Visualizations of t...

Pith tools

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