Pith. sign in

REVIEW 4 major objections 6 minor 11 cited by

Voyager: Long-Range and World-Consistent Video Diffusion for Explorable 3D Scene Generation

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Voyager proposes a video diffusion model that generates RGB and metric-depth frames together along a user-specified camera path, making 3D scene reconstruction a direct byproduct of generation.

desk verdict A genuinely new joint RGB-D video diffusion formulation with a solid system and ablations, but the headline metric-geometry claim rests on a self-referential evaluation loop that needs independent validation before publication. read the letter →

arxiv 2506.04225 v1 pith:DKDEQWLU submitted 2025-06-04 cs.CV

classification cs.CV
keywords videodiffusionRGB-Dgeneration3Dscenecameratrajectorycontrolworldconsistencypointcloudreconstructionlong-rangeautomaticdepthestimation
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

Voyager aims to make a single photograph the seed of an explorable 3D world. It claims that a video diffusion model can simultaneously generate color frames and metric depth frames along a user-defined camera trajectory, and that those RGB-D frames can be turned directly into a 3D point-cloud scene without any post-hoc reconstruction step such as structure-from-motion. If this works, it replaces the usual two-stage pipeline—generate video, then reconstruct geometry—with one end-to-end step, and it extends naturally to long or unbounded trajectories through an accumulating world cache. The payoff would be a practical way to turn arbitrary images into navigable virtual scenes for games, film, or simulation.

What carries the argument

The load-bearing mechanism is the world cache: a growing 3D point cloud, initialized by unprojecting the input image with an estimated depth map, that is projected into each target camera view to render partial RGB and partial depth conditions. The diffusion model conditions on these partial maps, generates aligned RGB-D frames, and the new frames are unprojected back into the cache, closing a loop; a normal-check point culling step removes redundant points while preserving geometric information. Supporting machinery includes concatenating RGB and depth along the height axis with a separator row, lightweight Control-blocks injected at each Transformer layer to strengthen geometric control, and overlapping smooth-sampling for autoregressive clip stitching. Together these components are what let the model keep long-range spatial consistency and output geometry that can be used for reconstruction directly.

What would settle it

Feed Voyager a single frame plus a long trajectory from a photorealistic synthetic scene with known camera poses and ground-truth metric depth, then compare each generated depth map to the renderer's depth under the true pose; if absolute-scale error grows with trajectory length or depth boundaries near occlusions disagree with the ground truth by more than the reported margin, the end-to-end metric world reconstruction claim is not holding.

Watch

Extended reading notes

Core claim

The paper proposes Voyager, a latent video diffusion model trained to output RGB and depth videos jointly for a specified camera path. The central claim is that depth-fused conditioning—rendering partial depth maps from a growing point cloud alongside partial RGB—removes the viewpoint inconsistency and occlusion hallucination that plague RGB-only warping conditions, and that generating depth simultaneously with color gives metric, directly usable geometry. The authors argue, to the best of their knowledge, that this is the first video model to jointly generate RGB and depth sequences with given camera trajectories, and that the design supports direct 3D reconstruction, long-range autoregressive extension via a point-culled world cache, and smooth clip transitions via overlapping sampling. Evidence comes from RealEstate10K view synthesis comparisons, Gaussian-splatting reconstruction from generated videos, and the WorldScore benchmark.

Load-bearing premise

The central claim depends on the assumption that the camera poses and metric depth estimates used as training and test labels are accurate enough that the model learns real scene geometry rather than reproducing the estimator's biases.

Editorial extensions

If this is right

  • If the model works as claimed, 3D scene reconstruction becomes a byproduct of video generation, removing structure-from-motion and multi-view stereo along with their artifacts.
  • Long or unbounded camera trajectories can be generated autoregressively while retaining global spatial consistency through the world cache.
  • Because depth is generated per-frame, the same model can act as a video depth estimator and support depth-consistent style transfer without retraining.
  • The data engine implies that visual quality can scale with diverse video sources rather than with manual 3D annotations.
  • End-to-end RGB-D generation makes image-to-3D creation from a single view feasible for scenes with multiple objects and correct spatial relationships.

Reading between the lines

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

  • Not claimed by the paper: the data engine could be closed into a self-improving loop, where Voyager's own generated depth is used to annotate new videos and expand the cache.
  • Not claimed by the paper: because depth-fused conditioning and Control-block injection are described as add-ons to a base video diffusion transformer, the same recipe should transfer to other video diffusion backbones.
  • Not claimed by the paper: the world cache is a persistent explicit 3D structure, so the framework could support interactive editing or relighting of the scene while generation continues, a use case the paper only gestures at.
  • Not claimed by the paper: evaluation on synthetic renders with ground-truth metric depth would quantify how much of the remaining error is absolute-scale drift versus occlusion-boundary uncertainty, since the paper's public metrics use estimated geometry.
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 / 6 minor

Summary. The manuscript proposes Voyager, a video-diffusion framework that, from a single input image and a user-specified camera trajectory, jointly generates aligned RGB and depth image sequences. The method builds on Hunyuan-Video by adding partial RGB-D conditioning rendered from an accumulating world cache, light-weight control blocks, point culling for memory efficiency, and auto-regressive smooth video sampling for long-range extension. A data engine automatically estimates camera poses and metric depth for over 100,000 video clips using VGGT, MoGE, and Metric3D, and the model is trained in three stages. The authors evaluate novel-view synthesis and Gaussian-splatting reconstruction on RealEstate10K, the WorldScore static benchmark, and ablations, claiming state-of-the-art visual and geometric performance and direct 3D reconstruction without structure-from-motion or multi-view stereo.

Significance. The paper makes a practically strong contribution if the geometric claims hold. It is, to my knowledge, the first video-diffusion system that jointly generates aligned RGB-D sequences for a user-specified camera path, and the world-cache with point culling plus smooth sampling provides a concrete mechanism for long-range world exploration. The data engine is a scalable way to obtain training pairs without manual 3D annotation, and the ablations in Table 4 and Figures 7–8 support the design choices of depth fusion and control blocks. The consistent gains over four open-source baselines on the RealEstate10K test subset and the top WorldScore score are encouraging. The significance is conditional, however, because the central geometric claim is currently supported only by consistency with the same estimator stack used to create the training labels; independent validation against ground-truth geometry is needed before the metric-reconstruction claim can be accepted.

major comments (4)
  1. [Sec. 5.1, Sec. 4.4, Eqs. (5)–(8)] The evaluation on RealEstate10K is self-referential in a load-bearing way. Section 5.1 states that for the test set 'we estimate the camera parameters and depth maps with the same pipeline in our data engine,' and Section 4.4 defines that pipeline (VGGT poses, MoGE depth, Metric3D rescaling) with Eqs. (5)–(8). This is exactly the pipeline that produces the training targets. Consequently Tables 1–2 quantify agreement with the estimator stack, not accuracy against true scene geometry, and any systematic scale or bias error in VGGT, MoGE, or Metric3D is invisible in the reported numbers. I request an evaluation on data with ground-truth cameras and depth (e.g., held-out Unreal Engine renders with GT depth/pose, or a real dataset with sensor depth), reporting absolute depth error and trajectory error without per-scene alignment, or with the alignment parameters explicitly disclosed.
  2. [Sec. 5.2, Table 2] The central claim of eliminating post-hoc 3D reconstruction is not exercised by the main geometric benchmark. Table 2 is titled 'Gaussian Splattig reconstruction' and Section 5.2 says Voyager's results are 'directly used in 3DGS reconstruction'; even the 'Voyager -' row appears to skip only the VGGT point-cloud initialization, not the 3DGS optimization itself. The claim that Voyager achieves 'end-to-end scene generation and reconstruction' should be validated by evaluating the generated RGB-D point cloud directly (e.g., depth accuracy, point-to-triangle distance, or rendered depth error against the reference) without a subsequent 3DGS fit, or the claim should be revised to make explicit that a lightweight optimization step remains.
  3. [Sec. 4.4, Eqs. (6)–(8)] The metric-depth alignment is too fragile to support the word 'metric' without further validation. Eq. (6) maps MoGE depth to Metric3D's 0.2–0.8 inter-quantile range, so a per-scene scale or bias error in Metric3D, or in the VGGT–MoGE least-squares alignment of Eq. (5), propagates directly into the training labels and into the test reference. Figure 12 compares Voyager point clouds only against VGGT, which is another estimator, not ground truth. Please report quantitative depth accuracy on a ground-truth benchmark and include a per-scene scale-error analysis.
  4. [Tables 1–3] No variance or significance information is reported. Most reported advantages are modest (e.g., PSNR 18.751 vs 18.278 for FlexWorld in Table 1), and with a 150-clip test subset these differences may be within noise. Please add bootstrap confidence intervals or paired significance tests, and report the number of independent scenes or clips used for each metric.
minor comments (6)
  1. [Table 2 caption] 'Gaussian Splattig' is a typo for 'Gaussian Splatting'.
  2. [Fig. 8 caption] The caption for Figure 8 is identical to that of Figure 7 and does not describe the point-culling and smooth-sampling ablations actually shown.
  3. [Fig. 8] The point counts ('num.: 194,253', etc.) are not explained in the text; please define how these counts are computed and state whether the claimed ~40% storage reduction is measured on these examples or on an average over clips.
  4. [Sec. 5.3] The statement that 'the camera movement in our results are larger than other methods' is made without quantitative support; please add a measure of camera displacement or rephrase it as a qualitative observation.
  5. [Dataset naming] The paper uses 'RealEstate' where the established name is RealEstate10K; please use the full name consistently.
  6. [Reproducibility] No code or data release is mentioned. Given the novelty claim and the dependence on a proprietary annotation pipeline, a statement on availability and ideally a release of the annotation script or model weights would substantially improve reproducibility.

Circularity Check

2 steps flagged · score 4.0 of 10

RealEstate geometry evaluation is self-referential: test cameras and depths are produced by the same VGGT/MoGE/Metric3D data engine used to create training targets, so Tables 1-2 evidence estimator agreement rather than independently verified metric geometry; the diffusion architecture itself is not circular.

  1. fitted input called prediction [Section 5.1, 'Dataset and Metrics']
    "Since the video clips do not provide ground-truth cameras, we estimate the camera parameters and depth maps with the same pipeline in our data engine."

    Training labels for Voyager's depth and camera conditioning are generated by the data engine of Sec. 4.4/Supplement C: VGGT poses, MoGE depth least-squares aligned to VGGT disparity (Eq. 5), then Metric3D-quantile rescaling (Eqs. 6-8). The RealEstate10K test geometry is then produced by this exact same pipeline because the dataset has no ground truth. Therefore the geometric rows of Tables 1-2, especially 'Voyager -' with generated depth, measure how well the model imitates the estimator stack it was trained on; any systematic scale or bias error in VGGT/MoGE/Metric3D appears in both the target and the reference and is undetectable. This is a self-referential validation of the metric-depth claim, not an independent test of metric correctness.

  2. self definitional [Section 4.4, 'Data Annotation' (Eqs. 6-8)]
    "Since our UE data provides metric depth values, we need to align all the estimated depth to a standard scale. We estimate the metric depth range of the scene using Metric3D [Hu et al. 2024] and map the previous depths into this range."

    For the real-video training data, 'metric depth' is not anchored to measured 3D geometry; it is defined by matching Metric3D's 0.2-0.8 quantile range (Eqs. 6-8), with the camera translation also scaled by the same factor. Because the same Metric3D-derived scale constructs the RealEstate test reference, the reported metric accuracy is by construction consistency with Metric3D rather than accuracy against an independent metric ground truth. A model that perfectly imitates the estimator would score perfectly while having arbitrarily wrong absolute scale. UE ground-truth depth in the third training stage partially mitigates this, but the in-domain RealEstate quantitative claims rest on this Metric3D-defined label/reference loop.

full rationale

Voyager is an empirical diffusion-systems paper rather than a first-principles derivation, so the imported-uniqueness, ansatz-by-citation, and renaming patterns do not apply. The core generation architecture—geometry-injected partial depth conditioning, joint RGB-D latent generation, world cache with point culling, and ControlNet-style blocks—is a trained neural system with no fitted constant that mathematically forces its outputs. The WorldScore benchmark is external, and the third training stage uses UE ground-truth depth, giving the central claim independent content. The significant issue is confined to the RealEstate10K geometry evaluation: Sec. 5.1 explicitly reuses the Sec. 4.4 data engine to create test cameras and depths, and Eqs. (5)-(8) show the metric scale is defined by least-squares alignment to VGGT/MoGE disparity and Metric3D quantile rescaling. Thus Tables 1-2 cannot distinguish genuine metric geometry from successful imitation of the training estimator; any systematic estimator bias is invisible. This is a self-referential validation loop, but it is not a case of the model's output being forced by construction, since the network could in principle generalize better or worse than its targets. Score 4 reflects that the architecture claim remains independent while the headline metric-depth accuracy on the in-domain benchmark is supported only by estimator self-consistency. No code or data is released, so the loop cannot be checked externally.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim sits on a stack of pretrained models and hand-set thresholds. The only fitted numbers are per-video scale and bias in the data engine, which are not global model parameters. The largest unverified assumption is that the VGGT/MoGE/Metric3D pipeline produces metric depth accurate enough to serve as both training target and test ground truth.

free parameters (4)
  • point culling normal angle threshold = 90 degrees
    Hand-chosen threshold to decide whether to keep points visible from the current view; reported to reduce stored points by about 40% (Sec. 4.3).
  • smooth sampling overlap ratio = half of one segment
    Overlap length of adjacent segments is set to half a segment; a design choice rather than a tuned value (Sec. 4.3).
  • metric depth quantile levels = 0.2 and 0.8
    Quantile levels in Eq. (6) used to map relative depth from MoGE into the metric range from Metric3D; chosen as robust percentiles.
  • depth alignment scale and bias per video = solved per video
    Least-squares scale and bias in Eq. (5) align MoGE disparity to VGGT disparity; fitted independently for each training video.
assumptions (3)
  • domain assumption HunyuanVideo latent diffusion is a suitable base model
    The entire architecture is built on HunyuanVideo; if this base model is flawed or unavailable, Voyager's results may not transfer.
  • domain assumption VGGT, MoGE, and Metric3D provide sufficiently accurate depth and camera poses for training data
    The data engine relies on these pretrained estimators to produce training targets; no ground-truth verification of their accuracy on the curated videos is provided.
  • domain assumption Point clouds from unprojected depth are an adequate persistent world representation
    The world cache stores points from unprojected depth, assuming depth maps are geometrically consistent across views, which is the property the method aims to achieve.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Voyager: Long-Range and World-Consistent Video Diffusion for Explorable 3D Scene Generation." pith.science (2026). https://pith.science/paper/DKDEQWLU

@misc{pith2026250604225,
  author       = {Pith},
  title        = {Pith review of: Voyager: Long-Range and World-Consistent Video Diffusion for Explorable 3D Scene Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DKDEQWLU}},
  note         = {Machine review of arXiv:2506.04225}
}
read the original abstract

Real-world applications like video gaming and virtual reality often demand the ability to model 3D scenes that users can explore along custom camera trajectories. While significant progress has been made in generating 3D objects from text or images, creating long-range, 3D-consistent, explorable 3D scenes remains a complex and challenging problem. In this work, we present Voyager, a novel video diffusion framework that generates world-consistent 3D point-cloud sequences from a single image with user-defined camera path. Unlike existing approaches, Voyager achieves end-to-end scene generation and reconstruction with inherent consistency across frames, eliminating the need for 3D reconstruction pipelines (e.g., structure-from-motion or multi-view stereo). Our method integrates three key components: 1) World-Consistent Video Diffusion: A unified architecture that jointly generates aligned RGB and depth video sequences, conditioned on existing world observation to ensure global coherence 2) Long-Range World Exploration: An efficient world cache with point culling and an auto-regressive inference with smooth video sampling for iterative scene extension with context-aware consistency, and 3) Scalable Data Engine: A video reconstruction pipeline that automates camera pose estimation and metric depth prediction for arbitrary videos, enabling large-scale, diverse training data curation without manual 3D annotations. Collectively, these designs result in a clear improvement over existing methods in visual quality and geometric accuracy, with versatile applications.

Figures

Figures reproduced from arXiv: 2506.04225 by the authors.

Figure 1
Figure 1. Voyager is a world-consistent video generation and reconstruction framework. Up: Voyager can generate 3D-consistent scene videos for world exploration following custom camera trajectories. Bottom: Voyager jointly generates aligned depth and RGB video for effective and direct 3D reconstruction. Real-world applications like video gaming and virtual reality often demand the ability to model 3D scenes that users can exp… view at source ↗
Figure 2
Figure 2. Partial RGB images and partial depth maps rendered from point [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview of Voyager: Given the input image and camera trajectories, we first render partial RGB images and depth maps for each viewpoint as the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Qualitative results on video generation. Compared to the baselines, our model can generate a more reasonable unseen region and meanwhile preserve [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Qualitative results on Gaussian Splatting reconstruction. Our results present much more details than the compared baselines. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Applications: (a) Long-range video generation. (b) Image-to-3D generation. (c) World-consistent video style transfer. (d) Monocular video depth [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 8
Figure 8. Figure 8: Qualitative results on ablation study. We compare the video models [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 7
Figure 7. Figure 7: Qualitative results on ablation study. We compare the video models [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 10
Figure 10. Figure 10: Overview of our scalable video data engine. [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: Training warp images compare. poses. To further improve the depth estimation, we then employ MoGE [Wang et al. 2024a] as a robust depth estimator. Specifically, we first convert the depth into disparity, and then use a least squares￾based optimization strategy to mini…
Figure 12
Figure 12. Figure 12: Comparison of initialization point clouds of ours and VGGT. [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 13
Figure 13. Figure 13: More Results [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]
Figure 14
Figure 14. Figure 14: More Visualization Results [PITH_FULL_IMAGE:figures/full_fig_p014_14.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 11 Pith papers

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

  1. CORGI: Consistency-Aware 3D Dog Reconstruction from a Single Image in the Wild

    cs.CV 2026-07 unverdicted novelty 7.0 of 10

    CORGI reconstructs high-fidelity, animatable 3D dogs from a single in-the-wild image via canonical orbital generation, deformable 3DGS anchored to D-SMAL, and self-supervised generative repair, without 3D supervision.

  2. 4DHumanDiff: Direct Text-to-4DGS Generation for Consistent 360-Degree Dynamic Humans

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A diffusion model trained on 60,000 fitted 4D Gaussian Splatting human clips generates text-prompted, view-consistent dynamic humans directly in 4D, over 10x faster than video-first pipelines.

  3. TSHA: A Benchmark for Visual Language Models in Trustworthy Safety Hazard Assessment Scenarios

    cs.CV 2026-03 unverdicted novelty 6.0 of 10

    TSHA is a new 80,000-pair benchmark for indoor safety hazard assessment; current vision-language models score roughly 45-85, and fine-tuning on TSHA raised Qwen2.5-VL-3B by 18.3 points on TSHA's test set.

  4. SceneExpander: Text-Guided 3D Scene Expansion via Free-Form View Insertion

    cs.CV 2026-03 conditional novelty 6.0 of 10

    A test-time adaptation method integrates a 3D-misaligned, AI-generated inserted view into a reconstructed 3D scene, preserving the captured region while extending it.

  5. CustomX: Unified Character, Action, and Scene Customization in Video World Models

    cs.CV 2025-12 conditional novelty 6.0 of 10

    AniX generates controllable videos of a user-supplied character performing typed actions inside a user-supplied 3D scene by fine-tuning a pre-trained video generator on small locomotion datasets.

  6. VDAWorld: World Modelling via VLM-Directed Abstraction and Simulation

    cs.CV 2025-12 conditional novelty 6.0 of 10

    A vision-language model writes a simulation program—grounded by segmentation and 3D tools—that predicts physically plausible futures from an image and caption, outperforming video generators on modified benchmarks.

  7. GeoWorld: Providing Full-frame Geometry Features to Facilitate 3D Scene Generation

    cs.CV 2025-11 conditional novelty 6.0 of 10

    GeoWorld improves image-to-3D scene generation by conditioning a video-diffusion model on full-frame geometry features extracted by a multi-view geometry model, yielding higher PSNR/SSIM/LPIPS than prior methods.

  8. VIGOR: VIdeo Geometry-Oriented Reward for Temporal Generative Alignment

    cs.CV 2026-03 conditional novelty 5.5 of 10

    A VGGT-based pointwise reprojection reward with geometry-aware sampling improves video geometric consistency via SFT/DPO and causal test-time search.

  9. WorldMirror: Universal 3D World Reconstruction with Any-Prior Prompting

    cs.CV 2025-10 conditional novelty 5.0 of 10

    A unified feed-forward network that ingests optional geometric priors and jointly predicts point maps, depth, camera poses, normals, and 3D Gaussians, reporting state-of-the-art results on multiple 3D benchmarks.

  10. 3D and 4D World Modeling: A Survey

    cs.CV 2025-09 conditional novelty 5.0 of 10

    A survey that defines 3D/4D world modeling, organizes methods into VideoGen, OccGen, and LiDARGen categories, and compiles datasets, metrics, and benchmark numbers.

  11. HunyuanWorld 1.0: Generating Immersive, Explorable, and Interactive 3D Worlds from Words or Pixels

    cs.CV 2025-07 conditional novelty 5.0 of 10

    A staged pipeline generates layered, mesh-based 3D worlds from text or images by combining panoramic diffusion, semantic layer decomposition, and video-based expansion.

Reference graph

Works this paper leans on

4 extracted references · 4 linked inside Pith · cited by 11 Pith papers

  1. [4]

    A mountain lake reflecting the peaks and a rustic cabin on stilts

    Dinet: Deformation inpainting network for realistic face visually dubbing on high resolution video. In Proceedings of the AAAI conference on artificial intelligence , Vol. 37. 3543–3551. Zibo Zhao, Zeqiang Lai, Qingxiang Lin, Yunfei Zhao, Haolin Liu, Shuhui Yang, Yifei Feng, Mingxin Yang, Sheng Zhang, Xianghui Yang, et al . 2025. Hunyuan3d 2.0: Scaling di...

  2. [2022]

    In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10684–10695. Runway. 2024. Introducing gen-3 alpha: A new frontier for video gneration. (2024). https://runwayml.com/research/introducing-gen-3-alpha Katja Schwarz, Denys Rozumnyi, Samuel Rota Bulò, Lorenzo P...

  3. [2023]

    arXiv preprint arXiv:2305.18264 (2023)

    Gen-l-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264 (2023). Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. 2025. Vggt: Visual geometry grounded transformer. arXiv preprint arXiv:2503.11651 (2025). Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xia...

  4. [2024]

    In European Conference on Computer Vision

    Citygaussian: Real-time high-quality large-scale scene rendering with gaus- sians. In European Conference on Computer Vision . Springer, 265–282. Yu Lu, Yuanzhi Liang, Linchao Zhu, and Yi Yang. 2024. Freelong: Training-free long video generation with spectralblend temporal attention. arXiv preprint arXiv:2407.19918 (2024). Baorui Ma, Huachen Gao, Haoge De...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.