Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

ArbiViewGen: Controllable Arbitrary Viewpoint Camera Data Generation for Autonomous Driving via Stable Diffusion Models

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read ArbiViewGen claims the first controllable arbitrary-view camera image generator for driving that needs only six camera images and poses, with no ground-truth images at novel viewpoints.

desk verdict Plausible self-supervised pipeline, but the quantitative evaluation is circular and the rotation-only homography assumption undermines near-field extrapolation; worth reviewing, not accepting as is. read the letter →

arxiv 2508.05236 v1 pith:CSF3PKJC submitted 2025-08-07 cs.CV

Yatong Lan , Jingfeng Chen , Yiru Wang , Lei He This is my paper
classification cs.CV
keywords arbitraryviewpointgenerationself-supervisedlearningdiffusionmodelsmulti-camerastitchingnovelviewsynthesisautonomousdrivingcross-viewconsistencynuScenes
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 tries to solve a bottleneck in autonomous driving data reuse: generating images from arbitrary camera viewpoints when no ground-truth image exists at that viewpoint. It claims a fully self-supervised framework, ArbiViewGen, that trains a latent diffusion model using only six surround cameras and their poses. Pseudo novel views are first stitched from the real images by a geometry-and-feature pipeline (FAVS); then a diffusion model is asked to reconstruct the original real images from those stitched views (CVC-SSL), so cross-view consistency is learned without external supervision. If correct, this would let one vehicle's collected data be re-rendered into the camera configuration of a different vehicle, reducing recollecting and re-annotation. The paper also proposes a quantitative evaluation by projecting colored LiDAR points into unseen views as sparse ground truth.

What carries the argument

The load-bearing mechanism is the FAVS-to-CVC-SSL training loop. FAVS builds pseudo-ground-truth novel views with the pure-rotation homography $H_{\mathrm{geom}}=K_2R_2R_1^{-1}K_1^{-1}$, refined by SIFT/FLANN/RANSAC and DBSCAN object alignment; CVC-SSL then trains a latent diffusion model with geometry-guided cross-view attention to reconstruct the real views from pseudo views sampled on both sides of each camera. The self-supervised reconstruction error replaces novel-view ground truth.

What would settle it

Take a calibrated six-camera rig with LiDAR, select a scene with a pedestrian within a few meters, stitch a pseudo view at a 30-degree extrapolated pose with FAVS, and compare the stitched image with a LiDAR-projected color image at the same pose: if close-range objects show parallax displacement, the rotation-only assumption is violated and the self-supervision target is corrupted.

Watch

Extended reading notes

Core claim

The central claim is that the missing ground truth for extrapolated viewpoints can be replaced by a closed self-supervised loop. ArbiViewGen stitches the six input images into a pseudo image at a sampled novel pose using a rotation-only homography refined by feature matching and object alignment (FAVS), then trains a shared-weight multi-branch latent diffusion U-Net with correspondence-aware attention to map the left and right pseudo views back to the original real view. Because the reconstruction target is a real image, no novel-view ground truth is required. The paper presents this as the first method for controllable arbitrary-view camera generation across multiple vehicle configurations

Load-bearing premise

The pipeline leans on the assumption that the mapping between cameras can be treated as a pure rotation, so the translation between cameras and the depth of objects can be ignored when stitching pseudo novel views.

Editorial extensions

If this is right

  • A pretrained ArbiViewGen could synthesize a new vehicle's camera views directly from an existing vehicle's six-camera data, reusing driving logs across platforms.
  • Because training needs only images and poses, fleets with camera-only rigs could participate in data generation without LiDAR or depth sensors.
  • The LiDAR-projected sparse ground-truth evaluation offers a way to score extrapolated views quantitatively, not just qualitatively.
  • The same cyclic-reconstruction idea could be applied to other multi-view tasks where target-view supervision is missing.

Reading between the lines

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

  • Editorial extension: the rotation-only homography in FAVS ignores the translation term $K_2(tn^\top/d)K_1^{-1}$, so at close range the pseudo ground truth misaligns; adding a cheap monocular depth prior inside FAVS is the natural next test.
  • Editorial extension: the self-supervision measures consistency with stitched inputs, not fidelity to the true scene; in disoccluded or textureless regions the model could learn plausible but incorrect content that sparse LiDAR evaluation cannot catch.
  • Editorial extension: the same loop could be extended to time, using temporally adjacent keyframes as pseudo target views, which might capture dynamic object motion without explicit depth.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. ArbiViewGen proposes a diffusion-based framework for generating camera images at arbitrary viewpoints from a six-camera surround-view setup, with the aim of reusing driving data across vehicle configurations. The method has two central components: FAVS, which stitches source camera images into pseudo novel views using a rotation-only homography refined by SIFT/RANSAC, clustering-based object alignment, and adaptive fusion; and CVC-SSL, which trains a latent diffusion model to reconstruct the original camera images from those pseudo views, thereby providing self-supervised training without ground truth at extrapolated poses. Evaluation is performed on nuScenes by projecting colored LiDAR point clouds into target views and computing sparse-pixel PSNR/SSIM/MAE/RMSE, with comparisons against a single baseline, DriveSuprim.

Significance. If the approach is sound, it addresses a practically important problem: enabling cross-platform reuse of multi-camera driving data without requiring depth sensors or ground-truth novel-view images. The self-supervised training loop is conceptually appealing, and the paper is honest about the lack of ground-truth extrapolated views. The method also ships a concrete evaluation protocol for sparse novel-view metrics, which is a useful contribution. However, the current evidence does not yet establish the central claim. The evaluation is at least partly circular because the sparse reference pixels are colored from the same input cameras used to construct the training pseudo-views, and the geometric backbone of the pseudo-views is based on a far-field approximation that is questionable for the near-field objects that dominate driving scenes. The manuscript would be strengthened by independent novel-view validation, a depth-aware or parallax-aware geometric model, and a broader set of baselines. As it stands, the central claims are plausible but under-supported.

major comments (3)
  1. [Experiment: Dataset, Evaluation, Table 1] The quantitative evidence is not independent of the training data. The sparse reference pixels are produced by coloring LiDAR points with colors sampled from the six input cameras and projecting them into target views (Experiment, Dataset). The model is trained to reconstruct those same real camera images from FAVS pseudo-views (Algorithm 1). Consequently, high PSNR/SSIM values in Table 1 largely measure reprojection self-consistency, not correctness of truly unseen content. The metrics are also computed only at sparse projected points; dense errors in extrapolated regions are not penalized. I recommend an evaluation against independent novel-view data (e.g., held-out cameras or temporally adjacent frames) and/or dense geometric consistency metrics before the claimed 'first ... controllable arbitrary view' result can be accepted.
  2. [Methods, Stage 1: Geometric Transformation Foundation] The pseudo-ground-truth geometry is built on H_geom = K2 R2 R1^-1 K1^-1, which drops the translation term present in H_planar = K2(R - t n^T/d) K1^-1. The text justifies this with a far-field approximation, but nuScenes contains pedestrians and vehicles at 5-20 m from cameras with baselines on the order of 1-2 m; for such points t n^T/d is not negligible and can cause tens-of-pixel misalignments. The SIFT/RANSAC and object-alignment steps correct only sparse keypoints and cannot remove depth-dependent parallax in the densely warped image. Since CVC-SSL trains the diffusion model to reconstruct real views from these geometrically distorted pseudo-views, the model may learn to hallucinate content that hides the homography error. This is exactly the regime where novel-view extrapolation is hardest. The conclusion's admission of 'limitations in capturing fine-grained structural details in hi
  3. [Experiment: Novel View Evaluation, Table 1] Only one baseline, DriveSuprim, is reported. DriveSuprim is a trajectory-selection/planning method adapted with rotation augmentation; it is not a representative novel-view synthesis or multi-view diffusion baseline. Methods discussed in Related Work (MVDiffusion, StreetCrafter, DriveX, 3DGS variants) are not quantitatively compared, nor is any ablation isolating the effect of pose extrapolation distance/angle. Without these comparisons and an analysis of failure modes as a function of baseline/translation, the claim that ArbiViewGen is 'the first method capable of controllable arbitrary view camera image generation' is under-supported. Please add at least one reconstruction-based and one diffusion-based baseline with the same sparse-projection protocol, and report metrics separately for small and large extrapolations.
minor comments (4)
  1. [Throughout] Inconsistent typography: 'Arbiviewgen' in the abstract versus 'ArbiViewGen' elsewhere; 'FA VS' has a stray space; 'synthesis' is misspelled in the Methods overview. Please standardize.
  2. [Algorithm 1: CVC-SSL] The number K of pseudo target poses and the sampling distribution for P_p-left and P_p-right are never specified. Please state K, the sampling range, and how left/right offsets are selected for reproducibility.
  3. [Figure 2 and qualitative results] Captions mention rotations (e.g., 27.5°, 35°) but not translations. Since the method claims arbitrary viewpoints, clarify the full pose perturbation used (angle and position offset) in the qualitative and quantitative experiments.
  4. [Experiment: Metrics] Table 1 reports sparse-pixel PSNR/SSIM/MAE/RMSE but does not state the number of sparse pixels per image or how points are colored and occluded. This information is essential for interpreting the metric values and for reproducing the evaluation.

Circularity Check

2 steps flagged · score 6.0 of 10

Partial circularity: pseudo-novel training views are generated from the same six images used as reconstruction targets, and the LiDAR-projected evaluation reference is colored from those same images, so training and quantitative validation reduce to self-consistency with the input views.

  1. self definitional [Methods, Overview and CVC-SSL Self-Supervised Training Process (Algorithm 1)]
    "Apply the FA VS algorithm to synthesize pseudo images at the sampled poses: I_p ← FAVS({I_1,...,I_6},{P_1,...,P_6}, P_p) ... For each real image, use the pseudo images from both sides (left and right) as input to the diffusion model. The model, equipped with geometry-guided cross-view attention, learns to reconstruct the original real image as its prediction target."

    FAVS is applied to the full set of six real images, which includes the same I_i that the model is trained to reconstruct. Each training input therefore already contains a warped copy of the target, and the learned mapping is the inverse of FAVS's own stitching/warp operation. The pseudo-novel views are not independent of the target; the 'prediction' of the original view is equivalent by construction to undoing a FAVS transform of a set containing that view. At inference, novel views are still conditioned on FAVS pseudo-views, so the extrapolated output is a refinement of a deterministic function of the input images rather than an independent novel-view prediction.

  2. other [Experiment, Dataset (evaluation benchmark construction)]
    "By coloring LiDAR point clouds using the original six cameras and projecting them to target viewpoints to generate sparse reference points, we construct a quantitative evaluation benchmark for extrapolated viewpoints."

    The sparse 'ground-truth' pixels at extrapolated viewpoints are obtained by coloring LiDAR with the original six cameras and projecting with the same calibration. The generated images are also synthesized from those six cameras. Thus an algorithm that merely warps or copies input pixels will match this reference by construction, and the metric cannot detect geometric errors (e.g., the dropped translation term in H_geom) because both prediction and reference are derived from the same source images. The quantitative claim is therefore a self-consistency measure, not an independent validation of arbitrary-view generation.

full rationale

No self-citation chain is involved; the circularity is in the closed training/evaluation loop. The paper explicitly states that pseudo ground truth is constructed from real images and used as input while the real images serve as supervision, and that evaluation reference points are colored from the original six cameras. Because both the synthesis and the reference are functions of the same input images, the reported quantitative gains are partly forced by construction. The diffusion refinement and qualitative comparisons provide some independent content, so the score is 6 rather than 8-10.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central claim rests on a pure-rotation homography approximation, a self-referential LiDAR-projection evaluation, pretrained Stable Diffusion transfer, homography-based attention, and an untested generalization from nearby pseudo poses to arbitrary target poses. Six families of hand-set or fitted parameters are either unreported or determined by experiment, and no code is available to pin them down.

free parameters (6)
  • Loss weighting hyperparameters alpha_L and beta_L = alpha_L = 0.1, beta_L = 0.01
    Total loss L_total = L_main + alpha_L * L_geo + beta_L * L_perceptual, stated as 'determined through experiment' in Methods, Loss Function Design.
  • Object alignment strength beta = range [0,1], exact value unspecified
    Haligned uses beta to control object alignment in FAVS Stage 3; the value is never reported.
  • Feature matching thresholds and dynamic alpha for Hbase = minmatches, consistency threshold, alpha schedule unspecified
    Selection among Hfeature, blended, and Hgeometric depends on these thresholds in FAVS Stage 2; none are given.
  • Fusion weights parameters gamma, sigma, quality weights, primary weights = unspecified
    FAVS Stage 4 distance weight exponent gamma, gradient weight sigma, and global quality/primary weights are hand-chosen and unreported.
  • DBSCAN epsilon and min_samples = unspecified
    Object cluster detection in FAVS Stage 3 requires these parameters; exact values are absent.
  • Pseudo target pose sampling distribution = unspecified
    Algorithm 1 samples K pseudo poses to the left and right of each real camera; the offset distribution and K values are not reported.
assumptions (6)
  • domain assumption Pure-rotation homography approximates the true source-to-target transform, ignoring translation (H_geom ≈ K2 R2 R1^-1 K1^-1).
    Introduced in Methods Stage 1 under the far-field approximation. Near objects violate this and produce incorrect pseudo-ground-truth geometry.
  • domain assumption Colored LiDAR point clouds projected into target views provide valid sparse ground truth for evaluating novel-view quality.
    The evaluation benchmark colors LiDAR using the same six cameras that are inputs to the generator, making the metric self-referential.
  • domain assumption Pretrained Stable Diffusion priors transfer to driving scenes and pseudo stitched inputs.
    The architecture is initialized from Stable Diffusion UNet; no analysis of domain shift is provided.
  • domain assumption Cross-view correspondences can be parameterized by homography for attention masks and positional encodings.
    Methods, Geometry-Guided Cross-View Attention assumes p_j ~ H_ij p_i for all features; non-planar or depth-varying content is approximated.
  • standard math SIFT/FLANN/RANSAC matching plus DBSCAN clustering reliably identifies object-level correspondences across views.
    Classic CV tools, but the reliability criteria in FAVS Stages 2 and 3 are heuristic and their thresholds are not reported.
  • ad hoc to paper Reconstructing original real views from nearby FAVS pseudo views transfers to arbitrary novel poses at inference.
    CVC-SSL trains only on pseudo poses sampled beside real cameras; the paper provides no ablation or evidence that this inverse mapping generalizes to large or arbitrary extrapolations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ArbiViewGen: Controllable Arbitrary Viewpoint Camera Data Generation for Autonomous Driving via Stable Diffusion Models." pith.science (2026). https://pith.science/paper/CSF3PKJC

@misc{pith2026250805236,
  author       = {Pith},
  title        = {Pith review of: ArbiViewGen: Controllable Arbitrary Viewpoint Camera Data Generation for Autonomous Driving via Stable Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CSF3PKJC}},
  note         = {Machine review of arXiv:2508.05236}
}
read the original abstract

Arbitrary viewpoint image generation holds significant potential for autonomous driving, yet remains a challenging task due to the lack of ground-truth data for extrapolated views, which hampers the training of high-fidelity generative models. In this work, we propose Arbiviewgen, a novel diffusion-based framework for the generation of controllable camera images from arbitrary points of view. To address the absence of ground-truth data in unseen views, we introduce two key components: Feature-Aware Adaptive View Stitching (FAVS) and Cross-View Consistency Self-Supervised Learning (CVC-SSL). FAVS employs a hierarchical matching strategy that first establishes coarse geometric correspondences using camera poses, then performs fine-grained alignment through improved feature matching algorithms, and identifies high-confidence matching regions via clustering analysis. Building upon this, CVC-SSL adopts a self-supervised training paradigm where the model reconstructs the original camera views from the synthesized stitched images using a diffusion model, enforcing cross-view consistency without requiring supervision from extrapolated data. Our framework requires only multi-camera images and their associated poses for training, eliminating the need for additional sensors or depth maps. To our knowledge, Arbiviewgen is the first method capable of controllable arbitrary view camera image generation in multiple vehicle configurations.

Figures

Figures reproduced from arXiv: 2508.05236 by the authors.

Figure 1
Figure 1. Pipeline of ArbiViewGen for controllable arbitrary-view image generation. FAVS generates pseudo ground-truth views via geometry-guided feature stitching. CVC-SSL trains a latent diffusion model with cross-view consistency to generate multi-view images from arbitrary poses using only 6-camera inputs and pose information—without requiring ground-truth extrapolated views. generate high-quality pseudo ground truth for n… view at source ↗
Figure 2
Figure 2. Qualitative comparison of novel-view synthesis under 27.5°, 35°, -13.75° and -17.5° camera rotations. Row 1 shows the original 6-camera images from nuScenes (Caesar et al. 2020). Row 2 displays results from DriveSuprim (Yao et al. 2025) using rotation-based augmentation. Row 3 displays synthesized views utilizing geometric transformations. Row 4 presents pseudo-views generated by our FAVS module. Row 5 illustrates t… view at source ↗
Figure 3
Figure 3. Qualitative comparison across six camera views. Comparison shows that our method achieves better visual align￾ment and consistency than DriveSuprim, Geometric, and FAVS across different camera views (e.g., CAM BACK with rotate 14° and CAM FRONT with rotate 11°) [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison across six camera views. Comparison shows that our method achieves better visual align￾ment and consistency than DriveSuprim, Geometric, and FAVS across different camera views (e.g., CAM BACK with rotate -14° and CAM FRONT with rotate -11°) [PIT…
Figure 5
Figure 5. Figure 5: Qualitative comparison across six camera views. Comparison shows that our method achieves better visual align￾ment and consistency than DriveSuprim, Geometric, and FAVS across different camera views (e.g., CAM BACK with rotate 28° and CAM FRONT with rotate 22°) [PITH_…
Figure 6
Figure 6. Figure 6: Qualitative comparison across six camera views. Comparison shows that our method achieves better visual align￾ment and consistency than DriveSuprim, Geometric, and FAVS across different camera views (e.g., CAM BACK with rotate -28° and CAM FRONT with rotate -22°) [PIT…

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. A Survey on the Applications of Generative Artificial Intelligence in Automated Driving Systems Test Scenario Generation Methods

    cs.SE 2025-12 reject novelty 4.0 of 10

    A literature survey of scenario-generation methods for ADS testing that adds an unvalidated AII/RAS/OCS metric suite and ODD-difficulty schema, undermined by inconsistent calculations in the worked examples.

Reference graph

Works this paper leans on

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

  1. [1]

    H.; Vora, S.; Liong, V

    Caesar, H.; Bankiti, V.; Lang, A. H.; Vora, S.; Liong, V. E.; Xu, Q.; Krishnan, A.; Pan, Y.; Baldan, G.; and Beijbom, O. 2020. nuScenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 11621--11631

    work page 2020

  2. [2]

    Dhariwal, P.; and Nichol, A. Q. 2021. Diffusion models beat GANs on image synthesis. In Advances in Neural Information Processing Systems (NeurIPS)

    work page 2021

  3. [3]

    Guo, J.; Ding, Y.; Chen, X.; Chen, S.; Li, B.; Zou, Y.; Lyu, X.; Tan, F.; Qi, X.; Li, Z.; and Zhao, H. 2025. DiST‑4D: Disentangled Spatiotemporal Diffusion with Metric Depth for 4D Driving Scene Generation. arXiv preprint arXiv:2503.15208

    arXiv 2025

  4. [4]

    Huang, N.; Wei, X.; Zheng, W.; An, P.; Lu, M.; Zhan, W.; Tomizuka, M.; Keutzer, K.; and Zhang, S. 2024. S ^ 3 Gaussian: Self‑Supervised Street Gaussians for Autonomous Driving. arXiv preprint arXiv:2405.20323

    arXiv 2024

  5. [5]

    Hwang, S.; Kim, M.; Kang, T.; Kang, J.; and Choo, J. 2024. VEGS: View Extrapolation of Urban Scenes in 3D Gaussian Splatting using Learned Priors. arXiv preprint arXiv:2407.02945

    arXiv 2024

  6. [6]

    Karras, T.; Aittala, M.; Laine, S.; Herva, J.; and Lehtinen, J. 2022. Elucidating the design space of diffusion-based generative models. NeurIPS

    work page 2022

  7. [7]

    Kerbl, B.; Kopanas, G.; Leimk \"u hler, T.; and Drettakis, G. 2023. 3D Gaussian Splatting for Real‑Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4)

    work page 2023

  8. [8]

    P.; and Welling, M

    Kingma, D. P.; and Welling, M. 2013. Auto-encoding variational Bayes. arXiv preprint arXiv:1312.6114

    arXiv 2013

Show all 25 references
  1. [9]

    Lin, C.; Zhuang, B.; Sun, S.; Jiang, Z.; Cai, J.; and Chandraker, M. 2024. Drive‑1‑to‑3: Enriching Diffusion Priors for Novel View Synthesis of Real Vehicles. arXiv preprint arXiv:2412.14494v1. Submitted Dec 19, 2024; accessed Jul 2025

    2024 arXiv

  2. [10]

    V.; Tokmakov, P.; Zakharov, S.; and Vondrick, C

    Liu, R.; Wu, R.; Hoorick, B. V.; Tokmakov, P.; Zakharov, S.; and Vondrick, C. 2023 a . Zero‑1‑to‑3: Zero‑shot One Image to 3D Object. In ICCV, 9264--9275. IEEE

    2023

  3. [11]

    Liu, Y.; Lin, C.; Zeng, Z.; Long, X.; Liu, L.; Komura, T.; and Wang, W. 2023 b . SyncDreamer: Generating Multiview‑consistent Images from a Single‑view Image. arXiv preprint arXiv:2309.03453

    2023 arXiv

  4. [12]

    Lu, C.; Lin, Y.; Chen, Q.; Bao, J.; Li, D.; Zhang, W.; Yang, D.; Gu, S.; Yuan, L.; and Zhang, L. 2022. DPM-Solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. NeurIPS

    2022

  5. [13]

    Miao, S.; Huang, J.; Bai, D.; Yan, X.; Zhou, H.; Wang, Y.; Liu, B.; Geiger, A.; and Liao, Y. 2025. EVolSplat: Efficient Volume‑based Gaussian Splatting for Urban View Synthesis. arXiv preprint arXiv:2503.20168. CVPR2025

    2025 arXiv

  6. [14]

    P.; Tancik, M.; Barron, J

    Mildenhall, B.; Srinivasan, P. P.; Tancik, M.; Barron, J. T.; Ramamoorthi, R.; and Ng, R. 2020. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In Computer Vision – ECCV 2020. Springer International Publishing

    2020

  7. [15]

    Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Ommer, B. 2022 a . High-resolution image synthesis with latent diffusion models. CVPR

    2022

  8. [16]

    Rombach, R.; et al. 2022 b . Stable Diffusion. https://github.com/CompVis/stable-diffusion

    2022

  9. [17]

    Shi, X.; Chen, L.; Wei, P.; Wu, X.; Jiang, T.; Luo, Y.; and Xie, L. 2024. DHGS: Decoupled Hybrid Gaussian Splatting for Driving Scene. arXiv preprint arXiv:2407.16600. Cs.CV

    2024 arXiv

  10. [18]

    Song, Y.; Meng, C.; and Ermon, S. 2020. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502

    2020 arXiv

  11. [19]

    V.; and Chen, M

    Sun, S.; Zhao, C.; Sun, Z.; Chen, Y. V.; and Chen, M. 2024. SplatFlow: Self‑Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving. arXiv preprint arXiv:2411.15482. Cs.CV

    2024 arXiv

  12. [20]

    Tang, S.; Zhang, F.; Chen, J.; Wang, P.; and Furukawa, Y. 2023. MVDiffusion: Enabling Holistic Multi-view Image Generation with Correspondence‑Aware Diffusion. In NeurIPS 2023 (Spotlight)

    2023

  13. [21]

    N.; Kaiser, .; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, .; and Polosukhin, I. 2017. Attention is all you need. In NeurIPS

    2017

  14. [22]

    Wang, Q.; Fan, L.; Wang, Y.; Chen, Y.; and Zhang, Z. 2024. FreeVS: Generative View Synthesis on Free Driving Trajectory. arXiv preprint arXiv:2410.18079

    2024 arXiv

  15. [23]

    Yan, Y.; Xu, Z.; Lin, H.; Jin, H.; Guo, H.; Wang, Y.; Zhan, K.; Lang, X.; Bao, H.; Zhou, X.; and Peng, S. 2025. StreetCrafter: Street View Synthesis with Controllable Video Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2025

  16. [24]

    Yang, Z.; Pan, Z.; Yang, Y.; Zhu, X.; and Zhang, L. 2024. DriveX: Driving View Synthesis on Free‑form Trajectories with Generative Prior. arXiv preprint arXiv:2412.01717v2. Version 2 PDF, submitted Dec 2024; accessed Jul 2025

    2024 arXiv

  17. [25]

    M.; and Wu, Z

    Yao, W.; Li, Z.; Lan, S.; Wang, Z.; Sun, X.; Alvarez, J. M.; and Wu, Z. 2025. DriveSuprim: Towards Precise Trajectory Selection for End‑to‑End Planning. arXiv preprint arXiv:2506.06659. Submitted 7 June 2025; revised 22 June 2025

    2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.