Pith. sign in

REVIEW 5 major objections 6 minor 2 cited by

PanoWan: Lifting Diffusion Video Generation Models to 360{\deg} with Latitude/Longitude-aware Mechanisms

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read PanoWan turns text prompts into seamless 360-degree video.

desk verdict A practical panoramic-video recipe with genuine ablations, but the latitude-aware noise argument only matches moments and the actual distribution is non-Gaussian for the simplest interpolation, so the central prior-preserving claim is unsupported. read the letter →

arxiv 2505.22016 v2 pith:RYY36BPU submitted 2025-05-28 cs.CV

Yifei Xia , Shuchen Weng , Siqi Yang , Jingqi Liu , Chengxuan Zhu , Minggui Teng , Zijian Jia , Han Jiang
show 1 more author
Boxin Shi
This is my paper
classification cs.CV
keywords panoramicvideogenerationtext-to-videodiffusionequirectangularprojectionlatitude-awaresamplingrotatedsemanticdenoisingpaddedpixel-wisedecodingLoRAfine-tuning360-degreedataset
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

PanoWan's claim is that a conventional text-to-video diffusion model can be lifted to generate coherent 360-degree panoramic video, rather than building a panorama generator from scratch. The paper argues that three minimal, sphere-aware mechanisms—latitude-aware sampling, rotated semantic denoising, and padded pixel-wise decoding—close the gap between equirectangular and ordinary video representations, while the new PANOVID dataset supplies enough captioned 360-degree footage for the model to learn the lifted representations. If this is right, immersive 360-degree content for VR, games, and embodied-AI simulation can be produced from ordinary text prompts by fine-tuning only about 1.6% of a pre-trained model's parameters.

What carries the argument

The load-bearing object is a sphere-aware wrapper around a frozen flow-matching DiT video model. Latitude-aware sampling defines the initial latent noise as $P'(x,y) = \operatorname{sgn}(BI(P,x,y)) \sqrt{BI(P^2,x,y)}$, where the horizontal sampling coordinate is remapped by $\cos((2y+1-R)\pi/(2R))$ before bilinear interpolation; the paper proves this preserves the first two moments of the noise and shows by a 1-D DFT argument that its spectral support matches the physical frequency of the sphere at each latitude. Rotated semantic denoising applies a circular shift $s_t = t \bmod W$ to the latent before each denoising step and unrolls it after, which spreads accumulated transition error over all longitudes and suppresses seam artifacts by a factor approaching $1/W$. Padded pixel-wise decoding circularly extends the denoised latent by $r$ columns, decodes, and center-crops, giving the VAE decoder the cross-seam context it was never trained on. LoRA fine-tuning on roughly 1.6% of parameters keeps the original generative priors intact while adapting the backbone to these new coordinate conventions.

What would settle it

Take the latitude-aware sampled noise $P'$ on polar rows and test whether its empirical distribution is materially non-Gaussian, for example with a Kolmogorov-Smirnov test against $N(0,1)$; if the distribution is far from Gaussian and yet the generated polar regions stay undistorted, the claimed role of the noise remapping would be unsupported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the generative prior of Wan 2.1, a pre-trained text-to-video model, survives transfer to the panorama when the noise schedule, the denoising trajectory, and the VAE decoding are each made aware of spherical geometry. Latitude-aware sampling remaps initial noise so that horizontal frequency content follows the cos(latitude) shrinkage of equirectangular projection; rotated semantic denoising rolls the latent code by a step-dependent shift so that seam errors spread over all longitudes instead of accumulating at one boundary; padded pixel-wise decoding supplies the VAE decoder with circular context across the seam. Trained with LoRA on the 13K-clip, 944-hour PANOVID dataset, the resulting model reports state-of-the-art scores on both general video metrics and panorama-specific metrics, and it performs zero-shot super-resolution, semantic inpainting, video outpainting, and long-video generation without task-specific training.

Load-bearing premise

The load-bearing premise is that after the latitude remapping, the interpolated noise is close enough to Gaussian for the pre-trained diffusion model to denoise it well, even though the paper verifies only the mean and variance, not the full distribution.

Editorial extensions

If this is right

  • Adding latitude-aware noise initialization and rotated denoising to other text-to-video backbones is claimed to be sufficient to produce seamless 360-degree videos with minimal retraining.
  • The same trained model can be reused, without extra fine-tuning, for super-resolution, semantic inpainting, and video outpainting of user-provided panoramic videos.
  • Long panoramic videos can be produced by a latent-space windowing scheme that denoises temporally overlapping chunks and blends them linearly, though the paper notes content is still forgotten over long horizons.
  • The PANOVID dataset, with over 13K captioned clips and 944 hours of footage, is positioned as a sufficient data foundation for this lifting task, replacing the much smaller WEB360 collection used by prior methods.

Reading between the lines

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

  • The three mechanisms are largely training-time recipes, so a plausible next test is whether the same pipeline lifts other backbones, such as larger DiT video models, without re-tuning the shift schedule or padding width.
  • Because the seam-error argument averages over many denoising steps, the benefit of rotated denoising should scale with the number of steps; a short-schedule variant might retain less of the benefit, which is a testable prediction the paper does not make explicitly.
  • The paper's own limitation section acknowledges content forgetting in long videos, so the claimed robustness is bounded by the backbone's temporal memory; memory-aware caching or recurrent conditioning is the natural extension.
  • The appendix proof only verifies that the remapped noise has mean zero and unit variance; checking whether its full distribution is close to Gaussian at high latitudes would tell whether latitude-aware sampling works as advertised or only approximately.
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

5 major / 6 minor

Summary. The paper proposes PanoWan, a framework that adapts the pre-trained text-to-video diffusion model Wan 2.1 to 360-degree panoramic video generation. The method introduces three mechanisms: latitude-aware sampling to mitigate equirectangular-projection distortion at high latitudes, rotated semantic denoising to distribute seam errors across longitudes during iterative denoising, and padded pixel-wise decoding to provide the VAE decoder with cross-seam context. The authors also contribute PANOVID, a captioned panoramic video dataset assembled from existing 360-degree sources and filtered with a five-stage pipeline. The paper reports quantitative and qualitative comparisons against 360DVD and DynamicScaler, ablations for each proposed module, and zero-shot demonstrations for super-resolution, inpaint-based editing, video outpainting, and long-video generation.

Significance. If the central claims hold, the paper offers a practical, parameter-efficient recipe for lifting conventional video diffusion priors to the panoramic domain, fine-tuning only about 1.6% of the base model parameters. The dataset contribution is potentially valuable to the community, and the paper's per-module ablations give genuine evidence about where the gains come from. The strongest part of the paper is the engineering: the pipeline is clearly described, the LoRA-based fine-tuning is economical, and the downstream applications are plausible and well illustrated. However, the main theoretical bridge between conventional and panoramic diffusion---the latitude-aware sampling---is not rigorously established: the supplied proof only matches two moments of the noise distribution, while the resulting noise field is in fact strongly non-Gaussian. Since flow-matching training in Eq. (1) assumes Gaussian initial latents, this gap is load-bearing for the claim that the pre-trained generative prior is preserved. The frequency-domain argument in Eqs. (7)-(8) is also heuristic.

major comments (5)
  1. [Sec. 4.2 / Appendix 7.1] The assertion that latitude-aware sampling produces noise that 'approach[es] the distribution on which the diffusion models are pre-trained' is unsupported. Appendix 7.1 proves only that E[P'] = 0 and E[Var P'] = 1. For the simplest equal-weight bilinear case, P' = sign((Z1+Z2+Z3+Z4)/4) * (1/2) sqrt(Z1^2+Z2^2+Z3^2+Z4^2), which has marginal density f(p) = 4|p|^3 exp(-2p^2); this density vanishes at p = 0 and peaks near |p| = 0.87, differing markedly from N(0,1). Because Eq. (1) trains the denoiser with z0 ~ N(0,I), the inference-time initial latents are out of distribution. Please provide a quantitative distributional check (e.g., Kolmogorov-Smirnov or energy distance) or evidence that the denoiser is robust to this non-Gaussian noise; otherwise the central mechanism for preserving Wan 2.1's prior is not established.
  2. [Sec. 4.2, Eqs. (7)-(8)] The frequency-domain analysis is heuristic and internally inconsistent. Eq. (7) concludes max f_car,x(y) <= R f_max, but the next paragraph sets 'max f_car,x = 2R' and 'f_max = 2' without relating units or deriving the value. Eq. (8) states that the spectral support is 'reduced to' about 2R cos(theta) after warping, but this is not derived from the DFT of a warped, interpolated discrete signal. Please either provide a rigorous sampling-theoretic derivation or clearly label this passage as qualitative motivation rather than a proof.
  3. [Table 1 / Sec. 5.4] The ablation 'W/o LAS' cannot isolate the benefit of latitude-aware frequency alignment, because removing LAS also removes the distribution shift of the non-Gaussian remapped noise. The comparison therefore confounds frequency alignment with distribution shift. To support the mechanism, compare LAS against a control that warps while preserving Gaussianity (e.g., resampling with variance normalization only), or report the distributional distance of the noise field before and after remapping.
  4. [Sec. 3 / Sec. 7.3] The dataset statistics are inconsistent. The pipeline segments videos into 10-second clips, yet the paper reports 'over 13K video clips totaling approximately 944 hours,' which implies an average clip length of about 4.3 minutes. Please clarify whether the final collection contains longer merged clips or whether the hour count refers to source videos before segmentation. As written, the scale claim is off by more than an order of magnitude relative to 13K ten-second clips (about 36 hours).
  5. [Sec. 5.3 / Abstract] The state-of-the-art claim rests on comparisons with only two baselines (360DVD and DynamicScaler). PanoDiT is excluded because its code is unavailable, and VideoPanda is not evaluated. Since the abstract and conclusion claim state-of-the-art performance across seven metrics, please include published numbers from prior methods even when code is unavailable, or restrict the claim to the evaluated baselines.
minor comments (6)
  1. [Sec. 5.2] The 'End Continuity' metric is described only as 'the mean absolute pixel difference across the left and right boundaries'; please define the formula precisely, including which rows/columns are compared and whether the computation is in RGB or latent space.
  2. [Throughout] There are typographical inconsistencies: 'Frechét' should be 'Fréchet' and 'DynamicScalar' appears alongside 'DynamicScaler' in several places.
  3. [Sec. 4.1, Eq. (2)] The notation 'PERP' is introduced for the equirectangular projection but is never used afterward; consider removing it or using it for consistency.
  4. [Sec. 4.2, Eq. (5)] The interpolation operator Interp_P is not defined for out-of-range coordinates introduced by the horizontal remapping; please specify the boundary handling (e.g., circular padding or clamping).
  5. [Sec. 7.3, Fig. 8] The figure caption says 'before balancing the semantics,' but the main text describes selecting up to 200 clips per category; please clarify that the histogram corresponds to the stage before that selection step.
  6. [Sec. 4.3, Eqs. (9)-(10)] The uniform-permutation claim for the rotated denoising needs qualification: the statement that the logical coordinates approach a uniform permutation over T steps depends on the relationship between the step count T and the width W; please state this condition explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PanoWan's derivation chain is self-contained, and its ablations are genuine.

full rationale

I examined the paper for the six enumerated circularity patterns. The core claim is that PanoWan lifts a pre-trained text-to-video model to panorama via latitude-aware sampling, rotated semantic denoising, padded pixel-wise decoding, and the PANOVID dataset. No parameter in the method is fitted to the evaluation metrics and then renamed as a prediction: LAS is a fixed deterministic remapping of i.i.d. Gaussian noise, RSD is a fixed circular shift schedule, and PPD is a fixed padding/cropping operation. The ablations in Table 1 and Figure 4 compare the full model against variants without each module, which is genuine empirical evidence rather than a circular reduction. The Appendix proof shows only E[P']=0 and E[Var P']=1, and the paper's statement that this 'approach[es] the distribution on which the diffusion models are pre-trained' is a fragility/correctness concern, not a circularity: the construction is not defined in terms of the target result, and the moment conditions do not by themselves force the desired full Gaussian distribution. The 'End Continuity' metric is newly introduced by the authors, but it is a measurement choice, not an input to training or a fitted parameter. There are no load-bearing self-citations, no uniqueness theorem imported from the authors' prior work, and no known empirical pattern merely renamed. The comparison with only two prior methods is a completeness concern, not circularity. Therefore the paper shows no significant circularity.

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

The central claim rests on several assumptions and hand-picked values: the transferability of Wan2.1 priors, the assumed frequency-scaling relation for equirectangular projection, the acceptability of moment-matched but non-Gaussian noise, and the quality of the collected PANOVID dataset. The data-pipeline thresholds (motion, aesthetics, per-category cap) and the unspecified padding width r are free parameters that affect the results but are not justified by a fitted objective.

free parameters (4)
  • motion_score_threshold = 0.4
    In the PANOVID filtering pipeline, clips with normalized optical-flow-based motion scores below 0.4 are discarded (Sec 7.3). This hand-chosen threshold directly determines which videos enter the training set.
  • aesthetic_score_threshold = 3 (Q-Align)
    Clips with average aesthetic score below 3 are removed during filtering (Sec 7.3). This hand-selected threshold controls the minimum visual quality of training data.
  • per_category_cap = 200
    To balance semantic categories, at most 200 clips with the highest aesthetic scores are retained per POI category (Sec 7.3). This directly determines the final dataset composition and diversity.
  • padding_width_r = unspecified
    Padded pixel-wise decoding (Sec 4.3) uses r columns of circular context, but the value of r is never stated. It is a free parameter of the method that affects seam quality and computational cost.
assumptions (4)
  • domain assumption Equirectangular projection preserves the frequency relationship given in Eq. (4): f_car,x = R cos(theta) f_sph.
    The design of latitude-aware sampling is based on this scaling. The paper introduces it with 'Assuming that warping preserves content' (Sec 4.1), making it a modeling assumption rather than an exact theorem.
  • domain assumption Pre-trained text-to-video priors from Wan 2.1 can be effectively transferred to the panoramic domain by fine-tuning only 1.6% of parameters via LoRA.
    The whole method depends on this transferability. The paper does not compare to full fine-tuning or to other adaptation methods, so the effectiveness of this minimal fine-tuning is assumed.
  • ad hoc to paper A noise prior that preserves only the first two moments (zero mean and unit variance) is acceptable to a pre-trained diffusion model.
    The proof in Sec 7.1 shows E[P']=0 and E[Var P']=1, but the interpolated noise P' is not shown to be Gaussian. The paper asserts that this 'approach[es] the distribution on which the diffusion models are pre-trained' without testing whether the non-Gaussian character affects generation.
  • domain assumption The PANOVID dataset is of sufficient quality, diversity, and size for training text-to-panoramic-video models.
    The dataset is assembled from external sources with an automated filtering pipeline based on popularity, motion, aesthetics, and caption similarity. No human validation of the resulting captions or video quality is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PanoWan: Lifting Diffusion Video Generation Models to 360{\deg} with Latitude/Longitude-aware Mechanisms." pith.science (2026). https://pith.science/paper/RYY36BPU

@misc{pith2026250522016,
  author       = {Pith},
  title        = {Pith review of: PanoWan: Lifting Diffusion Video Generation Models to 360\deg with Latitude/Longitude-aware Mechanisms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RYY36BPU}},
  note         = {Machine review of arXiv:2505.22016}
}
read the original abstract

Panoramic video generation enables immersive 360{\deg} content creation, valuable in applications that demand scene-consistent world exploration. However, existing panoramic video generation models struggle to leverage pre-trained generative priors from conventional text-to-video models for high-quality and diverse panoramic videos generation, due to limited dataset scale and the gap in spatial feature representations. In this paper, we introduce PanoWan to effectively lift pre-trained text-to-video models to the panoramic domain, equipped with minimal modules. PanoWan employs latitude-aware sampling to avoid latitudinal distortion, while its rotated semantic denoising and padded pixel-wise decoding ensure seamless transitions at longitude boundaries. To provide sufficient panoramic videos for learning these lifted representations, we contribute PanoVid, a high-quality panoramic video dataset with captions and diverse scenarios. Consequently, PanoWan achieves state-of-the-art performance in panoramic video generation and demonstrates robustness for zero-shot downstream tasks. Our project page is available at https://panowan.variantconst.com.

Figures

Figures reproduced from arXiv: 2505.22016 by the authors.

Figure 1
Figure 1. PanoWan is a text-based panoramic video generation framework. It lifts pre-trained [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The pipeline of our proposed PanoWan, aware of spherical coordinates. To avoid latitudinal [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Visual comparison results with existing text-based panoramic video generation methods. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Qualitative evaluation of proposed latitude/longitude-aware mechanisms. (a) With the [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Additional comparison results with existing text-based panoramic video generation methods. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Additional application results, showcasing the zero-shot capabilities for downstream tasks. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Visualization of failure cases. regions and exhibits relatively limited scene consistency. In contrast, our PanoWan achieves the most coherent and visually consistent results across diverse scenarios. Additional application results. We provide additional examples acros…
Figure 8
Figure 8. Figure 8: Category distribution of PANOVID dataset before balancing the semantics. • Video segmentation. We segment each video into 10-second clips using the PySceneDetect library, ensuring that each clip contains a single continuous scene without abrupt transitions. • Vision-la…
Figure 9
Figure 9. Figure 9: Representative samples from the PANOVID dataset. References [1] Miraikan 360-degree video dataset. https://www.miraikan.jst.go.jp/en/research/ AccessibilityLab/dataset360/. [2] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, …

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. ABot-3DWorld 0: A Universal World Model to Explore Any 3D Space

    cs.CV 2026-07 unverdicted novelty 6.0 of 10

    A unified pipeline lifts any text/image/video input into a Spatial Generative Primitive, explores it with 3D-consistent panoramic video, and reconstructs photorealistic 3DGS worlds with stronger rich-input fidelity th...

  2. EmoSpace: Immersive Affective Image Generation Guided by Fine-Grained Emotion Prototypes

    cs.CV 2026-02 conditional novelty 6.0 of 10

    EmoSpace generates emotion-controlled images and VR panoramas via a dynamic bank of 1,024 CLIP-space emotion prototypes, reporting higher fine-grained emotional alignment than baseline diffusion models.

Reference graph

Works this paper leans on

41 extracted references · 16 canonical work pages · cited by 2 Pith papers

  1. [1]

    https://www.miraikan.jst.go.jp/en/research/ AccessibilityLab/dataset360/

    Miraikan 360-degree video dataset. https://www.miraikan.jst.go.jp/en/research/ AccessibilityLab/dataset360/

    work page

  2. [2]

    Qwen2.5-VL technical report.arXiv preprint arXiv:2502.13923, 2025

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2.5-VL technical report.arXiv preprint arXiv:2502.13923, 2025

    arXiv 2025

  3. [3]

    Is space-time attention all you need for video understanding? InICML, 2021

    Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? InICML, 2021

    2021

  4. [4]

    Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023

    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

    arXiv 2023

  5. [5]

    360+x: A panoptic multi-modal scene understanding dataset

    Hao Chen, Yuqi Hou, Chenyuan Qu, Irene Testini, Xiaohan Hong, and Jianbo Jiao. 360+x: A panoptic multi-modal scene understanding dataset. InProc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

    work page 2024

  6. [6]

    Geometry fidelity for spherical images

    Anders Christensen, Nooshin Mojab, Khushman Patel, Karan Ahuja, Zeynep Akata, Ole Winther, Mar Gonzalez-Franco, and Andrea Colaco. Geometry fidelity for spherical images. In Proc. of European Conference on Computer Vision, 2024

    work page 2024

  7. [7]

    Two-frame motion estimation based on polynomial expansion

    Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. InImage Analysis, 2003

    work page 2003

  8. [8]

    CameraCtrl II: Dynamic scene exploration via camera- controlled video diffusion models.arXiv preprint arXiv:2503.10592, 2025

    Hao He, Ceyuan Yang, Shanchuan Lin, Yinghao Xu, Meng Wei, Liangke Gui, Qi Zhao, Gordon Wetzstein, Lu Jiang, and Hongsheng Li. CameraCtrl II: Dynamic scene exploration via camera- controlled video diffusion models.arXiv preprint arXiv:2503.10592, 2025

    arXiv 2025

Show all 41 references
  1. [9]

    Latent video diffusion models for high-fidelity long video generation.arXiv preprint arXiv:2211.13221, 2022

    Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation.arXiv preprint arXiv:2211.13221, 2022

    2022 arXiv

  2. [10]

    Imagen video: High definition video generation with diffusion models.arXiv preprint arXiv:2210.02303, 2022

    Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models.arXiv preprint arXiv:2210.02303, 2022

    arXiv

  3. [11]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. InProc. of the International Conference on Learning Representations, 2022. 16

    2022

  4. [12]

    DynamicScaler: Seamless and scalable video generation for panoramic scenes

    Liu Jinxiu, Lin Shaoheng, Li Yinxiao, and Yang Ming-Hsuan. DynamicScaler: Seamless and scalable video generation for panoramic scenes. InProc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025

    2025

  5. [13]

    A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2020

    Diederik P Kingma, J Adam Ba, and J Adam. A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2020

    2020 arXiv

  6. [14]

    Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    2024 arXiv

  7. [15]

    Li, Jeremy N

    Benjamin J. Li, Jeremy N. Bailenson, Adam Pines, Walter J. Greenleaf, and Leanne M. Williams. A public database of immersive vr videos with corresponding ratings of arousal, valence, and correlations between head movements and self report measures.Frontiers in Psychology, 2017

    2017

  8. [16]

    4K4DGen: Panoramic 4D generation at 4K resolution.arXiv preprint arXiv:2406.13527, 2024

    Renjie Li, Panwang Pan, Bangbang Yang, Dejia Xu, Shijie Zhou, Xuanyang Zhang, Zeming Li, Achuta Kadambi, Zhangyang Wang, Zhengzhong Tu, et al. 4K4DGen: Panoramic 4D generation at 4K resolution.arXiv preprint arXiv:2406.13527, 2024

    2024 arXiv

  9. [17]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pag...

    2024

  10. [18]

    Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    2022 arXiv

  11. [19]

    GenEx: Generating an explorable world.Proc

    Taiming Lu, Tianmin Shu, Junfei Xiao, Luoxin Ye, Jiahao Wang, Cheng Peng, Chen Wei, Daniel Khashabi, Rama Chellappa, Alan Yuille, and Jieneng Chen. GenEx: Generating an explorable world.Proc. of the International Conference on Learning Representations, 2025

    2025

  12. [20]

    Be- yond the frame: Generating 360 ◦ panoramic videos from perspective videos.arXiv preprint arXiv:2504.07940, 2025

    Rundong Luo, Matthew Wallingford, Ali Farhadi, Noah Snavely, and Wei-Chiu Ma. Be- yond the frame: Generating 360 ◦ panoramic videos from perspective videos.arXiv preprint arXiv:2504.07940, 2025

    2025 arXiv

  13. [21]

    VidPanos: Generative panoramic videos from casual panning videos

    Jingwei Ma, Erika Lu, Roni Paiss, Shiran Zada, Aleksander Holynski, Tali Dekel, Brian Curless, Michael Rubinstein, and Forrester Cole. VidPanos: Generative panoramic videos from casual panning videos. InProc. of ACM SIGGRAPH Asia, 2024

    2024

  14. [22]

    SphereD- iff: Tuning-free omnidirectional panoramic image and video generation via spherical latent representation.arXiv preprint arXiv:2504.14396, 2025

    Minho Park, Taewoong Kang, Jooyeol Yun, Sungwon Hwang, and Jaegul Choo. SphereD- iff: Tuning-free omnidirectional panoramic image and video generation via spherical latent representation.arXiv preprint arXiv:2504.14396, 2025

    2025

  15. [23]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InICCV, 2023

    2023

  16. [24]

    GEN3C: 3D-informed world- consistent video generation with precise camera control

    Xuanchi Ren, Tianchang Shen, Jiahui Huang, Huan Ling, Yifan Lu, Merlin Nimier-David, Thomas Müller, Alexander Keller, Sanja Fidler, and Jun Gao. GEN3C: 3D-informed world- consistent video generation with precise camera control. InProc. of IEEE/CVF Conference on Computer Vision...

    2025

  17. [25]

    Seaweed-7B: Cost-effective training of video generation foundation model.arXiv preprint arXiv:2504.08685, 2025

    SeaweadTeam, Ceyuan Yang, Zhijie Lin, Yang Zhao, Shanchuan Lin, Zhibei Ma, Haoyuan Guo, Hao Chen, Lu Qi, Sen Wang, et al. Seaweed-7B: Cost-effective training of video generation foundation model.arXiv preprint arXiv:2504.08685, 2025

    2025 arXiv

  18. [26]

    Make-a-video: Text-to-video generation without text-video data

    Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. InICLR, 2023

    2023

  19. [27]

    Imagine360: Immersive 360 video generation from perspective anchor.arXiv preprint arXiv:2412.03552, 2024

    Jing Tan, Shuai Yang, Tong Wu, Jingwen He, Yuwei Guo, Ziwei Liu, and Dahua Lin. Imagine360: Immersive 360 video generation from perspective anchor.arXiv preprint arXiv:2412.03552, 2024

    2024 arXiv

  20. [28]

    Towards accurate generative models of video: A new metric & challenges

    Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 17

    2018 arXiv

  21. [29]

    Veo-Team, :, Agrim Gupta, Ali Razavi, Andeep Toor, Ankush Gupta, Dumitru Erhan, Eleni Shaw, Eric Lau, Frank Belletti, Gabe Barth-Maron, Gregory Shaw, Hakan Erdogan, Hakim Sidahmed, Henna Nandwani, Hernan Moraldo, Hyunjik Kim, Irina Blok, Jeff Donahue, José Lezama, Kory Mathews...

    2024

  22. [30]

    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. InProc. of Neural Information Processing...

    2024

  23. [31]

    VideoCLIP- XL: Advancing long description understanding for video clip models.arXiv preprint arXiv:2410.00741, 2024

    Jiapeng Wang, Chengyu Wang, Kunzhe Huang, Jun Huang, and Lianwen Jin. VideoCLIP- XL: Advancing long description understanding for video clip models.arXiv preprint arXiv:2410.00741, 2024

    2024 arXiv

  24. [32]

    360DVD: Controllable panorama video generation with 360-degree video diffusion model

    Qian Wang, Weiqi Li, Chong Mou, Xinhua Cheng, and Jian Zhang. 360DVD: Controllable panorama video generation with 360-degree video diffusion model. InProc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

    2024

  25. [33]

    Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    WanTeam, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    2025 arXiv

  26. [34]

    Q-Align: Teaching LMMs for visual scoring via discrete text-defined levels.arXiv preprint arXiv:2312.17090, 2023

    Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Chunyi Li, Liang Liao, Annan Wang, Erli Zhang, Wenxiu Sun, Qiong Yan, Xiongkuo Min, Guangtai Zhai, and Weisi Lin. Q-Align: Teaching LMMs for visual scoring via discrete text-defined levels.arXiv preprint arXiv:2312.17090, 2023

    2023 arXiv

  27. [35]

    WORLDMEM: Long-term consistent world simulation with memory.arXiv preprint arXiv:2504.12369, 2025

    Zeqi Xiao, Yushi Lan, Yifan Zhou, Wenqi Ouyang, Shuai Yang, Yanhong Zeng, and Xingang Pan. WORLDMEM: Long-term consistent world simulation with memory.arXiv preprint arXiv:2504.12369, 2025

    2025

  28. [36]

    VideoPanda: Video panoramic diffusion with multi-view attention.arXiv preprint arXiv:2504.11389, 2025

    Kevin Xie, Amirmojtaba Sabour, Jiahui Huang, Despoina Paschalidou, Greg Klar, Umar Iqbal, Sanja Fidler, and Xiaohui Zeng. VideoPanda: Video panoramic diffusion with multi-view attention.arXiv preprint arXiv:2504.11389, 2025

    2025 arXiv

  29. [37]

    Panonut360: A head and eye tracking dataset for panoramic video

    Yutong Xu, Junhao Du, Jiahe Wang, Yuwei Ning, Sihan Zhou, and Yang Cao. Panonut360: A head and eye tracking dataset for panoramic video. InProceedings of the 15th ACM Multimedia Systems Conference, 2024

    2024

  30. [38]

    CogVideox: Text-to-video diffusion models with an expert transformer

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. CogVideox: Text-to-video diffusion models with an expert transformer. InICLR, 2025

    2025

  31. [39]

    Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Vighnesh Birodkar, Agrim Gupta, Xiuye Gu, Alexander G

    Lijun Yu, José Lezama, Nitesh B. Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Vighnesh Birodkar, Agrim Gupta, Xiuye Gu, Alexander G. Hauptmann, Boqing Gong, Ming-Hsuan Yang, Irfan Essa, David A. Ross, and Lu Jiang. Language model beats diffusion – tokenize...

    2024

  32. [40]

    PanoDit: Panoramic videos generation with diffusion transformer

    Muyang Zhang, Yuzhi Chen, Rongtao Xu, Changwei Wang, JinMing Yang, Weiliang Meng, Jianwei Guo, Huihuang Zhao, and Xiaopeng Zhang. PanoDit: Panoramic videos generation with diffusion transformer. InProc. of the AAAI Conference on Artificial Intelligence, 2025

    2025

  33. [41]

    Holotime: Taming video diffusion models for panoramic 4d scene generation, 2025

    Haiyang Zhou, Wangbo Yu, Jiawen Guan, Xinhua Cheng, Yonghong Tian, and Li Yuan. Holotime: Taming video diffusion models for panoramic 4d scene generation, 2025. URL https://arxiv.org/abs/2504.21650. 18

    2025 arXiv

Pith tools

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