Pith. sign in

REVIEW 4 major objections 5 minor 39 references

RGE-GS: Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors

T0 review · 4 major / 5 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 RGE-GS claims that gating diffusion-generated pixels by a learned confidence map lets 3D Gaussian Splatting reconstruct never-scanned driving lanes without accumulating artifacts.

desk verdict RGE-GS is a sensible diffusion-plus-GS reconstruction recipe with a clean two-part idea, but its Para-Lane numbers are hard to trust because the diffusion prior was fine-tuned on the same dataset used for evaluation. read the letter →

arxiv 2506.22800 v3 pith:G7Q7AGZL submitted 2025-06-28 cs.CV

Sicong Du , Jiarun Liu , Qifeng Chen , Hao-Xiang Chen , Tai-Jiang Mu , Sheng Yang This is my paper
classification cs.CV
keywords 3DGaussianSplattingdiffusionpriorsreward-guidedreconstructionnovelviewsynthesisdrivingsceneexpansionconfidencemapdifferentiatedtrainingautonomoussimulation
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

RGE-GS tries to establish that diffusion priors can be reliably used to expand a single-pass driving reconstruction into unseen lanes, without accumulating artifacts, by gating every generated pixel through a learned confidence map before it enters 3D Gaussian Splatting. A second mechanism splits Gaussian optimization into missing, immature, and mature primitives so that only unfinished regions are trained on synthetic views. If correct, this would make sensor simulators for autonomous driving able to synthesize physically consistent views of lanes that were never scanned. The paper supports the claim with comparisons on Para-Lane and Waymo, reporting higher fidelity than prior diffusion-aided reconstruction baselines.

What carries the argument

The reward network is a lightweight U-Net that takes a generated image and outputs a pixel-wise confidence map between 0 and 1, trained jointly with the Gaussian scene using a reprojection error loss against colored LiDAR projections and a binarization regularization that keeps predictions decisive. The differentiated training strategy classifies Gaussian primitives into missing, immature, and mature: missing Gaussians are initialized from LiDAR or depth estimates, immature Gaussians receive extra optimization against the generated views, and mature Gaussians are frozen to preserve stability. The confidence map multiplies the reconstruction loss for generated views, so low-confidence pixels contribute less to training.

What would settle it

Run the official or a fully faithful implementation of the strongest compared method on the same four Para-Lane scenes and check whether RGE-GS still holds the 19.72 PSNR one-lane margin; separately, fine-tune the diffusion prior on a disjoint set of lanes and see whether the reconstruction advantage survives.

Watch

Extended reading notes

Core claim

The paper claims that diffusion-generated novel views can be safely folded into a 3D Gaussian Splatting reconstruction of a driving scene if each generated pixel is first weighted by a learned confidence score and the scene's Gaussians are then trained in differentiated stages. Applied to two base Gaussian splatters, this reward-guided integration yields higher PSNR, SSIM, and lower FID than the compared diffusion-aided reconstruction methods for one- and two-lane expansion on Para-Lane, and improves FID on Waymo especially at larger lateral shifts.

Load-bearing premise

The headline result assumes the re-implemented baseline comparisons run the published methods at their full intended strength, and that the diffusion prior fine-tuned on Para-Lane does not memorize the specific lanes used for evaluation.

Editorial extensions

If this is right

  • On Para-Lane one-lane expansion, RGE on Scaffold-GS reaches 19.72 PSNR and 0.586 SSIM, above the 19.07 and 0.537 of the strongest reimplemented baseline.
  • Ablations show removing the reward network drops PSNR from 19.72 to 19.24, and removing differentiated training raises training time from 89 to 112 minutes.
  • The reward module also improves a stronger video-diffusion backbone: RGE with stable video diffusion reaches 19.95 PSNR one-lane, above the 19.72 with the image-diffusion backbone.
  • On Waymo, RGE on Scaffold-GS keeps FID lowest at all lateral shifts from 1 to 5 meters, with the gap widening at larger shifts.
  • Because the reward network and training strategy are adapter modules, the same recipe can be placed on top of other Gaussian reconstruction frameworks beyond the two tested bases.

Reading between the lines

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

  • The pixel-wise reward map could be reused as an uncertainty signal for active scanning, telling a vehicle where to drive next to fill under-reconstructed regions.
  • Since the paper describes the reward network as lightweight and jointly trained, the same module could be bolted onto any diffusion-aided Gaussian pipeline, not just the two bases tested.
  • The paper fine-tunes its diffusion model on data built from the same Para-Lane dataset used for evaluation; a held-out-lane or cross-dataset evaluation would clarify whether the confidence map is filtering genuine geometric inconsistency or memorized content.
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 proposes RGE-GS, a framework for expanding driving-scene reconstructions by integrating diffusion-generated cross-lane views into 3D Gaussian Splatting. Two components are introduced: a reward network that produces pixel-wise confidence maps for generated views, and a differentiated training strategy that classifies Gaussians into missing, immature, and mature groups and applies different optimization treatments. Evaluations on Para-Lane report consistent gains over 3DGS, 2DGS, Street Gaussians, Scaffold-GS, and reimplemented FreeSim/ReconDreamer baselines, with additional FID results on Waymo and ablations supporting both components.

Significance. If the reported results are robust, the work is significant: it addresses a real problem in diffusion-aided reconstruction, namely the selective integration of generated views that may be physically inconsistent, and it demonstrates a modular component (reward-guided weighting) that can be combined with existing Gaussian frameworks. The paper ships no code yet but promises public release, and the ablations (Table 3) and two-dataset evaluation (Tables 1-2) are useful evidence. The main contributions are clearly stated and the method is described in sufficient detail for reproduction in principle. However, the central SOTA claim is currently not verifiable because of the missing training/evaluation split for the diffusion prior and because the baselines are reimplemented with a modified pipeline.

major comments (4)
  1. [§4.1, Table 1] The diffusion model is fine-tuned on a training set constructed from the Para-Lane dataset (~60,000 pairs), while the main quantitative evaluation is on four Para-Lane scenes. No scene-level split is reported. If the evaluation scenes are included in the diffusion fine-tuning data, the generated cross-lane views I_e used in Eq. (6) could be near-memorized samples from the evaluation distribution, so the PSNR/SSIM/LPIPS gains in Table 1 would not isolate the reward network or differentiated training. The Waymo FID results are distribution-level and use no cross-lane ground truth, so they do not resolve the overlap risk. Please report the exact scene IDs used for diffusion training versus evaluation, or re-run Table 1 with a disjoint scene split; without this, the headline state-of-the-art claim is unverifiable.
  2. [§4.1, Tables 1-2] The paper compares against FreeSim* and ReconDreamer*, which are described as reimplementations that 'directly use generated images of cross-lane views for training, bypassing the progressive view generation approach.' This is a substantive modification of the published methods, and it is not self-evident that it yields the 'upper-bound performance' claimed; the original progressive generation may be important for achieving the baselines' reported quality. A state-of-the-art claim against modified baselines is therefore not established. Please use official checkpoints or published numbers where available, or present a controlled comparison in which the only difference between RGE-GS and the baseline is the proposed module, and justify why the reimplementation cannot disadvantage the baselines.
  3. [§3.3, Eq. (5)] The reward network is described as being trained 'in an adversarial strategy,' but the loss in Eq. (5) contains no adversarial term: L_reproj is a regression loss against a LiDAR point-cloud projection and L_reg is a binarization regularizer. There is no discriminator, minimax objective, or game-theoretic component. This matters because the unsupervised confidence map is the central mechanism claimed to reject biased diffusion outputs; with the stated losses, the network could learn any score that is monotone in reprojection error, and the paper provides no analysis of what prevents degenerate solutions. Please remove the term 'adversarial,' formally define what makes the verification adversarial, or add an explicit adversarial component.
  4. [Tables 1-3] All Para-Lane conclusions are based on only four scenes, and no error bars, per-scene breakdowns, or significance tests are reported. Some of the claimed improvements are small in absolute terms (e.g., LPIPS 0.415 vs 0.424 in Table 1; LPIPS 0.415 vs 0.427 in Table 3), so random scene selection could change the qualitative conclusions. The Waymo FID results also lack confidence intervals. Please report per-scene results with standard deviations and the number of scenes, and, where possible, significance tests for the main comparisons.
minor comments (5)
  1. [§4.1] The heading contains a typo: 'Evalution' should be 'Evaluation,' and 'datset' in the first paragraph should be 'dataset.'
  2. [§4.1] The term 'upper-bound performance' is ambiguous and potentially misleading; consider replacing it with a precise statement about how the reimplementation relates to the original methods.
  3. [§3.4] The gradient threshold lambda = 5e-4 for identifying immature Gaussians and the choice of a fixed threshold are not analyzed; a short sensitivity study would strengthen the differentiated-training contribution.
  4. [Tables 1-2] The naming 'RGE on Scaffold-GS' is used in Table 2 but 'RGE on Scaffold-GS' is also the full method in Table 1; consider using one consistent notation throughout.
  5. [References] References [26] and [27] are both U-Net papers by the same authors with overlapping content; one may be redundant, and the citation style should be unified.

Circularity Check

1 steps flagged · score 6.0 of 10

Diffusion prior is fine-tuned on Para-Lane and evaluated on Para-Lane without a reported split; Table 1's SOTA margins may measure training-set recall rather than generative extrapolation.

  1. fitted input called prediction [Sec. 4.1 (diffusion model training) & Sec. 4.2 (Table 1)]
    "The training dataset is constructed from the Para-Lane dataset using method in [28], containing about 60,000 pairs. ... On the Para-Lane dataset, we select four scenes containing three lanes (left, middle, right), comprising 250 front-view images of each lane scanned at 1280×720@10Hz to evaluate the aforementioned method with publicly available baselines"

    The diffusion model is fine-tuned for 50k steps on a training set built from Para-Lane (Sec. 4.1), and the headline quantitative evaluation (Table 1) is done on four Para-Lane scenes. Since no scene-level split is reported, the evaluation scenes may be (indeed, with only four scenes and ~60k pairs, almost certainly are) present in the diffusion training pairs. The generated cross-lane images I_e, fed into the reconstruction loss Eq. (6), are then near-memorized ground-truth views rather than extrapolated novel views. Consequently, the PSNR/SSIM gains in Table 1 measure how well the diffusion prior memorized the test scenes (and how faithfully the Gaussian pipeline can copy that memory), not the reward network or differentiated training.

full rationale

The paper's central SOTA claim on Para-Lane is compromised by an unreported train/evaluation split. The diffusion prior is trained on ~60k Para-Lane pairs, then the same dataset's four scenes are used for evaluation. Without an explicit exclusion of these scenes from the training construction, the generated cross-lane images can be close to memorized ground-truth views, so the improvement in Table 1 is not attributable to the proposed reward guidance. The Waymo FID results (Table 2) are more independent because the diffusion model was not trained on Waymo and no ground-truth cross-lane images exist there; however, they are distribution-level metrics and do not rescue the central qualitative SOTA statement. No other self-referential loops were found: the reward network is supervised by LiDAR projections, the differentiated training is a standard adaptive optimization, and baselines are externally cited. The absence of a documented split is the key missing support that converts a potential empirical claim into a circular-looking one.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

No new physical entities are introduced. The main added components are the reward network and the differentiated training schedule, both of which rest on hand-set hyperparameters and domain assumptions. The fine-tuned diffusion prior is trained on the same dataset used for evaluation, which is a circularity/leakage concern handled in red_flags.

free parameters (3)
  • Immature Gaussian gradient threshold = 5e-4 (Section 3.4) and 4e-4 (Section 4.1)
    Hand-chosen threshold to classify Gaussians as immature vs mature; the paper gives two values without explanation, and the method's convergence behavior depends on it.
  • Loss weights = lambda_reproj=0.5, lambda_reg=0.3, lambda_RGB=0.8, lambda_Ie=0.01, lambda_Io=0.8
    Hand-selected weights balancing reprojection, binarization, RGB, SSIM, and LPIPS terms; they affect the learned reward map and reconstruction quality.
  • Depth scale calibration = not reported
    DepthAnything-V2 estimates are scaled by minimizing Euclidean distance to LiDAR; the scale factor is fit per scene and not reported.
assumptions (5)
  • domain assumption 3D Gaussian Splatting optimization converges to a high-quality scene under standard hyperparameters
    The method relies on 3DGS and Scaffold-GS as base reconstructions without re-deriving their properties.
  • domain assumption Colored LiDAR point cloud projection provides sufficient coarse supervision for reward map learning
    L_reproj in Eq. 3 assumes the projected colored LiDAR points align with the diffusion image, but LiDAR is sparse and may be absent in under-observed regions.
  • domain assumption Gradient magnitude is a reliable indicator of Gaussian maturity
    The differentiated training strategy classifies Gaussians by view-space positional gradient thresholds; no evidence is given that this correlates with true reconstruction completeness.
  • domain assumption DepthAnything-V2 depth can be scaled to metric using LiDAR
    Missing Gaussians are initialized from scaled monocular depth; if scaling is inaccurate, initialization may be biased.
  • ad hoc to paper Reward network's unsupervised objective yields meaningful confidence without collapsing to all-zero
    The bias constraint prevents all-zero output, but no analysis shows the learned C_e is not an arbitrary mask; validation is only indirect via MAE.
invented entities (1)
  • Reward network F_c independent evidence
    purpose: Predicts pixel-wise confidence C_e for each diffusion-generated image to weight its contribution during Gaussian reconstruction
    The paper validates the reward map through MAE against ground truth in Fig. 5 and ablations in Tab. 3; still a learned module rather than a physical entity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RGE-GS: Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors." pith.science (2026). https://pith.science/paper/G7Q7AGZL

@misc{pith2026250622800,
  author       = {Pith},
  title        = {Pith review of: RGE-GS: Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G7Q7AGZL}},
  note         = {Machine review of arXiv:2506.22800}
}
read the original abstract

A single-pass driving clip frequently results in incomplete scanning of the road structure, making reconstructed scene expanding a critical requirement for sensor simulators to effectively regress driving actions. Although contemporary 3D Gaussian Splatting (3DGS) techniques achieve remarkable reconstruction quality, their direct extension through the integration of diffusion priors often introduces cumulative physical inconsistencies and compromises training efficiency. To address these limitations, we present RGE-GS, a novel expansive reconstruction framework that synergizes diffusion-based generation with reward-guided Gaussian integration. The RGE-GS framework incorporates two key innovations: First, we propose a reward network that learns to identify and prioritize consistently generated patterns prior to reconstruction phases, thereby enabling selective retention of diffusion outputs for spatial stability. Second, during the reconstruction process, we devise a differentiated training strategy that automatically adjust Gaussian optimization progress according to scene converge metrics, which achieving better convergence than baseline methods. Extensive evaluations of publicly available datasets demonstrate that RGE-GS achieves state-of-the-art performance in reconstruction quality. Our source-code will be made publicly available at https://github.com/CN-ADLab/RGE-GS.

Figures

Figures reproduced from arXiv: 2506.22800 by the authors.

Figure 1
Figure 1. The workflow of our proposed RGE-GS for expansive scene reconstruction. We use a single trajectory clip to reconstruct gaussians [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The illustration of reward network. Network structure. As shown in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Comparisons for cross-lane view synthesis with publicly available methods, demonstrating our effectiveness of using a reward [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Ablations of cross-lane view synthesis results with different experiment settings. Notice the restoration of vegetation color and [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Visualization of reward map and layered diffusion output. The darker the areas in the reward map, the lower the confidence. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 34 canonical work pages

  1. [1]

    Patchmatch: A randomized correspon- dence algorithm for structural image editing

    Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspon- dence algorithm for structural image editing. ACM TOG, 28 (3):24, 2009. 2

    work page 2009

  2. [2]

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

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv:2311.15127, 2023. 1, 3, 7

    arXiv 2023

  3. [3]

    Gaussianpro: 3d gaussian splatting with progressive propagation

    Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. InICML, pages 1–18, 2024. 2

    work page 2024

  4. [4]

    Streetscapes: Large- scale consistent street view generation using autoregressive video diffusion

    Boyang Deng, Richard Tucker, Zhengqi Li, Leonidas Guibas, Noah Snavely, and Gordon Wetzstein. Streetscapes: Large- scale consistent street view generation using autoregressive video diffusion. In ACM SIGGRAPH , pages 27:1–27:11,

    work page

  5. [5]

    FreeSim: Toward free-viewpoint camera simu- lation in driving scenes

    Lue Fan, Hao Zhang, Qitai Wang, Hongsheng Li, and Zhaox- iang Zhang. FreeSim: Toward free-viewpoint camera simu- lation in driving scenes. InCVPR, pages 12004–12014, 2025. 1, 2, 5

    work page 2025

  6. [6]

    InstantSplat: Unbounded sparse-view pose-free gaus- sian splatting in 40 seconds

    Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, Zhangyang Wang, and Yue Wang. InstantSplat: Unbounded sparse-view pose-free gaus- sian splatting in 40 seconds. arXiv:2403.20309, 2024. 2

    arXiv 2024

  7. [7]

    MagicDrive: Street view generation with diverse 3D geometry control

    Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. MagicDrive: Street view generation with diverse 3D geometry control. InICLR, pages 1–20, 2024. 1, 2

    work page 2024

  8. [8]

    Magicdrive3d: Controllable 3d generation for any-view rendering in street scenes

    Ruiyuan Gao, Kai Chen, Zhihao Li, Lanqing Hong, Zhenguo Li, and Qiang Xu. Magicdrive3d: Controllable 3d generation for any-view rendering in street scenes. InICLR, pages 1–17,

    work page

Show all 39 references
  1. [9]

    Vista: A generalizable driving world model with high fidelity and versatile controllability

    Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. In NeurIPS, 2025. 2

    2025

  2. [10]

    Ggs: Generalizable gaussian splatting for lane switching in autonomous driving

    Huasong Han, Kaixuan Zhou, Xiaoxiao Long, Yusen Wang, and Chunxia Xiao. Ggs: Generalizable gaussian splatting for lane switching in autonomous driving. In AAAI, pages 3329–3337, 2025. 2

    2025

  3. [11]

    2D gaussian splatting for geometrically accu- rate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D gaussian splatting for geometrically accu- rate radiance fields. In ACM SIGGRAPH, pages 32:1–32:11,

  4. [12]

    Gs-planner: A gaussian-splatting- based planning framework for active high-fidelity reconstruc- tion

    Rui Jin, Yuman Gao, Yingjian Wang, Yuze Wu, Haojian Lu, Chao Xu, and Fei Gao. Gs-planner: A gaussian-splatting- based planning framework for active high-fidelity reconstruc- tion. In IEEE IROS, pages 11202–11209, 2024. 2

    2024

  5. [13]

    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):139:1–139:14,

  6. [14]

    Auto-encoding varia- tional bayes

    Diederik P Kingma and Max Welling. Auto-encoding varia- tional bayes. In ICLR, pages 1–14, 2014. 5

    2014

  7. [15]

    DrivingDiffu- sion: Layout-guided multi-view driving scene video genera- tion with latent diffusion model

    Xiaofan Li, Yifu Zhang, and Xiaoqing Ye. DrivingDiffu- sion: Layout-guided multi-view driving scene video genera- tion with latent diffusion model. In ECCV, pages 469–485,

  8. [16]

    Choose your sim- ulator wisely: A review on open-source simulators for au- tonomous driving

    Yueyuan Li, Wei Yuan, Songan Zhang, Weihao Yan, Qiyuan Shen, Chunxiang Wang, and Ming Yang. Choose your sim- ulator wisely: A review on open-source simulators for au- tonomous driving. IEEE TIV, 9(5):4861–4876, 2024. 1

    2024

  9. [17]

    Re- conX: Reconstruct any scene from sparse views with video diffusion model

    Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Re- conX: Reconstruct any scene from sparse views with video diffusion model. In ICLR, pages 1–22, 2025. 2

    2025

  10. [18]

    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. In NeurIPS, pages 133305– 133327, 2024. 2

    2024

  11. [19]

    Scaffold-GS: Structured 3D gaus- sians for view-adaptive rendering

    Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-GS: Structured 3D gaus- sians for view-adaptive rendering. In CVPR, pages 20654– 20664, 2024. 1, 5

    2024

  12. [20]

    InfiniCube: Unbounded and controllable dynamic 3D driving scene generation with world-guided video models

    Yifan Lu, Xuanchi Ren, Jiawei Yang, Tianchang Shen, Zhangjie Wu, Jun Gao, Yue Wang, Siheng Chen, Mike Chen, Sanja Fidler, and Jiahui Huang. InfiniCube: Unbounded and controllable dynamic 3D driving scene generation with world-guided video models. arXiv:2412.03934, 2024. 1

    2024 arXiv

  13. [21]

    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 synthe- sis. Commun. ACM, 65(1):99–106, 2021. 2

    2021

  14. [22]

    Recon- Dreamer: Crafting world models for driving scene recon- struction via online restoration

    Chaojun Ni, Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Wenkang Qin, Guan Huang, Chen Liu, Yuyin Chen, Yida Wang, Xueyang Zhang, Yifei Zhan, Kun Zhan, Peng Jia, Xianpeng Lang, Xingang Wang, and Wenjun Mei. Recon- Dreamer: Crafting world models for driving scene recon- struction v...

  15. [23]

    Para-Lane: Multi-lane dataset registering par- allel scans for benchmarking novel view synthesis

    Ziqian Ni, Sicong Du, Zhenghua Hou, Chenming Wu, and Sheng Yang. Para-Lane: Multi-lane dataset registering par- allel scans for benchmarking novel view synthesis. In 3DV, pages 1–10, 2025. 2, 5

    2025

  16. [24]

    Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild

    Weining Ren, Zihan Zhu, Boyang Sun, Jiaqi Chen, Marc Pollefeys, and Songyou Peng. Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild. In CVPR, pages 8931–8940, 2024. 2

    2024

  17. [25]

    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, 5, 7

    2022

  18. [26]

    U-Net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234–241, 2015. 5

    2015

  19. [27]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234–241, 2015. 4

    2015

  20. [28]

    FreeVS: Generative view synthesis on free driving trajectory

    Qitai Wang, Lue Fan, Yuqi Wang, Yuntao Chen, and Zhaoxi- ang Zhang. FreeVS: Generative view synthesis on free driving trajectory. In ICLR, pages 1–18, 2025. 1, 2, 5

    2025

  21. [29]

    Drivedreamer: Towards real-world- drive world models for autonomous driving

    Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jia- gang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world- drive world models for autonomous driving. InECCV, pages 55–72, 2024. 1, 2

    2024

  22. [30]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004. 4

    2004

  23. [31]

    Street gaussians: Modeling dynamic urban scenes with gaussian splatting

    Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians: Modeling dynamic urban scenes with gaussian splatting. In ECCV, pages 156–173,

  24. [32]

    StreetCrafter: Street view synthesis with controllable video diffusion models

    Yunzhi Yan, Zhen Xu, Haotong Lin, Haian Jin, Haoyu Guo, Yida Wang, Kun Zhan, Xianpeng Lang, Hujun Bao, Xiaowei Zhou, and Sida Peng. StreetCrafter: Street view synthesis with controllable video diffusion models. In CVPR, pages 822–832, 2025. 2

    2025

  25. [33]

    Depth anything v2

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. NeurIPS, 37:21875–21911, 2024. 4

    2024

  26. [34]

    Sgd: Street view synthesis with gaussian splatting and diffusion prior

    Zhongrui Yu, Haoran Wang, Jinze Yang, Hanzhang Wang, Zeke Xie, Yunfeng Cai, Jiale Cao, Zhong Ji, and Mingming Sun. Sgd: Street view synthesis with gaussian splatting and diffusion prior. In WACV, pages 3812–3822, 2025. 2

    2025

  27. [35]

    MapGS: Generalizable pretraining and data augmentation for online mapping via novel view synthesis

    Hengyuan Zhang, David Paz, Yuliang Guo, Xinyu Huang, Henrik I Christensen, and Liu Ren. MapGS: Generalizable pretraining and data augmentation for online mapping via novel view synthesis. arXiv:2501.06660, 2025. 1

    2025 arXiv

  28. [36]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836–3847, 2023. 5

    2023

  29. [37]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586–595,

  30. [38]

    DriveDreamer4D: World models are effective data machines for 4D driving scene representation

    Guosheng Zhao, Chaojun Ni, Xiaofeng Wang, Zheng Zhu, Xueyang Zhang, Yida Wang, 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. In CVPR, pages 12015–120...

    2025

  31. [39]

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

    Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Xinze Chen, Guan Huang, Xiaoyi Bao, and Xingang Wang. Drivedreamer- 2: Llm-enhanced world models for diverse driving video generation. In AAAI, pages 10412–10420, 2025. 2

    2025

Pith tools

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