Pith. sign in

REVIEW 4 major objections 5 minor 38 references

The paper argues that a rank-8 LoRA adapter on a standard perspective-video diffusion model is enough to generate seamless 360° panoramic video, treating the projection change as an 8-degree-of-freedom style shift.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 17:05 UTC pith:5OMB2MNN

load-bearing objection A plausible LoRA-based recipe for panoramic video generation, but the theoretical rank-sufficiency claim is an upper-bound argument that doesn't hold and is contradicted by the paper's own rank-8 ablation. the 4 major comments →

arxiv 2509.11092 v1 pith:5OMB2MNN submitted 2025-09-14 cs.CV cs.AI

PanoLora: Bridging Perspective and Panoramic Video Generation with LoRA Adaptation

Zeyu Dong , Yuyang Yin , Yuqi Li , Eric Li , Hao-Xiang Guo , Yikai Wang This is my paper
classification cs.CV cs.AI
keywords panoramic video generationLoRAlow-rank adaptationvideo diffusion modelequirectangular projectionleft-right consistencydegrees of freedom360-degree video
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that generating 360° panoramic video can be reduced to a low-rank adaptation of an existing perspective-video generation model. The authors claim that the perspective-to-panoramic projection has 8 degrees of freedom, and prove that a LoRA update with rank at least 8 can span that solution space, so fine-tuning just the LoRA adapters on a pretrained video diffusion model should suffice. They fine-tune such a model on roughly 1,000 synthetic panoramic videos and report state-of-the-art left-right seam closure and motion magnitude. A sympathetic reader would care because it suggests a lightweight, architecture-free route to immersive 360° video generation, in contrast to prior work that adds heavy auxiliary networks or large-scale training.

Core claim

The central claim is that the transformation from a perspective video to a panoramic equirectangular video is a low-dimensional projection change with 8 degrees of freedom (horizontal and forward translation, yaw, focal lengths, principal-point offsets, and scale). The paper proves that the rank of the output change induced by a LoRA update is bounded by the LoRA rank, so any adaptation that must cover the 8-dimensional solution space requires rank ≥ 8. On this basis, the authors fine-tune a pretrained video diffusion model with LoRA adapters of rank 16 on ~1,000 rendered panoramic videos and report left-right consistency of 0.99 and motion magnitudes approaching ground-truth levels, while r

What carries the argument

The rank-degree-of-freedom inequality: for a LoRA update Δθ = AB with A∈R^{p×r}, B∈R^{r×1}, the resulting output variation δF = J_F(θ)·Δθ has rank at most r, because rank(XY) ≤ min(rank X, rank Y). Since the perspective-to-panoramic projection is argued to have 8 degrees of freedom, the LoRA rank must be at least 8 to cover that solution space. The paper also reduces the naive 11-DoF projection model to 8 DoF using camera-pose statistics from its own 447 training clips, dropping roll, pitch, and vertical shift as low-variance parameters.

Load-bearing premise

The central threshold rank ≥ 8 depends on the claim that the perspective-to-panoramic transformation has only 8 degrees of freedom, a claim the paper supports only with statistics from its own 447 training clips, where pitch, roll, and vertical shift happen to vary very little.

What would settle it

Fine-tune the same LoRA pipeline on panoramic videos with substantial pitch or roll variation (for example, aerial footage with camera tilt), keeping rank 8; if left-right consistency drops below ~0.9 or the 360° seam breaks, the 8-DoF simplification is violated and the rank threshold fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • LoRA with rank ≥ 8 on a pretrained perspective-video diffusion model can produce panoramic video with proper projection geometry, reaching 0.99 left-right consistency and motion magnitudes near ground truth at rank 16.
  • Ranks below 8 underfit: rank 5 yields left-right consistency of 0.74 with visible seam breaks, and rank 8 only partially recovers to 0.95, supporting the theoretical rank threshold.
  • The method needs only ~1,000 training videos and about 10 hours on 8 A100 GPUs, suggesting that 360° video generation can be achieved with lightweight fine-tuning rather than new architectures.
  • The placement of LoRA adapters affects motion type: attention-only adapters favor lateral/rotational alignment, linear-only adapters favor appearance/texture modulation, and combining both yields balanced depth-aware motion.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The rank-8 threshold is derived from the camera-motion statistics of the paper's own synthetic dataset, where pitch, roll, and vertical shift have very small variance; on real-world panoramic footage with strong pitch or roll, the effective degrees of freedom would rise and the rank requirement would likely increase.
  • The theoretical proof shows only that the LoRA update has the capacity to span the 8-dimensional space, not that it actually does so; the observed left-right consistency of 0.95 at rank 8, versus the perfect closure the theory suggests, hints that the practical rank requirement may be higher than the DoF count.
  • A natural testable extension is to apply the same rank ≥ degrees-of-freedom principle to other projection transitions, such as perspective-to-fisheye or perspective-to-cubemap, where the DoF count differs and the rank threshold would change predictably.
  • Even the best configuration still produces motion magnitudes around half the ground-truth values (front 4.02 vs 7.64), suggesting that LoRA adaptation captures projection geometry but may need extra capacity or conditioning to fully match real-world dynamics.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes PanoLora, which treats text-to-panoramic-video generation as a low-rank adaptation (LoRA) problem on a pretrained perspective-video diffusion model (Wan2.1-14B). The authors first argue, from camera-projection geometry and statistics of their own 447 training clips, that the perspective-to-panoramic transformation has 8 degrees of freedom. They then present a 'theoretical proof' that a LoRA rank r ≥ 8 suffices to span this transformation space. Empirically, they fine-tune Wan2.1 on ~1,000 synthetic UE panoramic videos and report improvements over 360DVD, 360DVD+, and GenEX on left-right consistency and per-view optical-flow motion magnitude. Ablations vary LoRA placement (attention vs. linear layers) and rank (5, 8, 16, 32), concluding that rank 16 is the best configuration.

Significance. The reformulation of panoramic video generation as a lightweight LoRA adaptation is practically appealing and the paper includes a reasonably extensive ablation over rank and injection points. If the theoretical claim were valid, it would provide a simple design rule for choosing LoRA rank. However, the theoretical argument is not sound: it derives only an upper bound on the rank of the output change and then asserts sufficiency. The paper's own experiments contradict the rank-8 threshold (rank 8 gives L-R consistency 0.95, rank 16 gives 0.99). The empirical evaluation is also narrow, relying on only two custom metrics and a small set of baselines on synthetic data. The central contribution—the theoretical justification—is therefore unsupported, and the empirical results, while suggestive, do not establish state-of-the-art performance.

major comments (4)
  1. [Section 3.3, single-layer case] The proof derives rank(δF) ≤ min(rank(J_F), rank(Δθ)) ≤ r, which is an upper bound on the rank of the output variation. It then claims that if r ≥ 8, 'the changeable space can be fully covered.' This is a non-sequitur: an upper bound does not imply that a rank-r LoRA can realize any desired 8-dimensional output change, nor that the specific perspective-to-panoramic weight update lies in the reachable subspace. Moreover, δF is defined as a vector in R^q, so rank(δF) is at most 1; the argument appears to conflate the rank of a matrix with the dimension of the output space. No lower bound or reachability argument is provided.
  2. [Section 3.3, multi-layer and nonlinear extension] In the multi-layer linear case, δF is the sum of per-layer Jacobian contributions. The claim that each layer with rank r ≥ 8 yields a total variation spanning a space of dimension at least 8 is not proved: the contributions can cancel or overlap, so the rank of the sum can be much smaller than the sum of ranks. The extension to nonlinear networks invokes local linearity, but local linearity does not convert an upper bound into a lower bound. Thus the sufficiency result remains unsupported.
  3. [Section 3.1, Table 1 and Section 4.4, Table 3] The 8-DoF simplification is derived from the authors' own training clips (Table 1), not from an independent analysis of the target distribution; pitch and roll are discarded because their standard deviations are small in this particular dataset. If test data contain substantial pitch or roll variation, the threshold fails. More importantly, the paper's own ablation contradicts the claimed threshold: Table 3 reports L-R consistency 0.95 at rank 8 and 0.99 at rank 16, with rank 5 at 0.74. If r ≥ 8 were sufficient, rank 8 should achieve near-perfect closure. The empirical results therefore do not support the central claim and instead suggest a monotonic improvement up to rank 16.
  4. [Section 4, Main Results and Evaluation] The state-of-the-art claim is based on comparison with only one dedicated panoramic baseline (360DVD), a LoRA-augmented variant (360DVD+), and a general video model (GenEX). The evaluation uses only two custom metrics (left-right seam cosine similarity and optical-flow magnitude) on synthetic UE-generated scenes. No standard video generation metrics (e.g., FVD, LPIPS, CLIP score) or user study are reported, and no real-world 360° video test set is used. This is insufficient to establish the claimed superior performance and generalizability.
minor comments (5)
  1. [Section 3.3, notation] The LoRA parameterization Δθ = AB with A∈R^{p×r}, B∈R^{r×1} makes Δθ a vector, not a low-rank matrix. This conflicts with the standard LoRA update ΔW = BA used in the preceding section. Please reconcile the notation and clarify what 'rank' means for a vector.
  2. [Section 3.1, Eq. (3)] The eight degrees of freedom are listed as fx, fy, cx, cy, s, tx, ty, θyaw, but the text refers to 'scale factors' (plural). Please define s precisely and explain how it relates to the intrinsic matrix parameters.
  3. [Figures 2 and 4] The captions and text do not consistently match panel labels; for example, the text refers to panels (d–e) in some places and (a)–(e) in others. Please align all panel references with the figures.
  4. [Throughout] There are several typos and formatting issues, e.g., 'mulit-layer' in Section 3.3, '360 ◦' spacing, and inconsistent use of 'equirectangular' vs. 'panoramic.' These should be corrected.
  5. [Section 4, Implementation Details] The paper states that LoRA is applied to 'spatial and temporal layers' in the Related Works, but the implementation only mentions QKVO projections and FFNs in attention blocks. Clarify whether temporal modules are adapted and if not, remove the claim.

Circularity Check

2 steps flagged

The LoRA rank threshold r≥8 is not derived externally: the 8-DoF count is an empirical simplification of the authors' own training clips, and the proof only shows an upper bound on rank, not that a rank-8 LoRA spans the mapping; the paper's own ablation shows rank 8 underperforms rank 16.

specific steps
  1. fitted input called prediction [Section 3.1, Task Definition, Eq. (3)]
    "Based on the above analysis, we empirically simplify the camera motion model, retaining only the horizontal shift (tx), forward shift (ty), yaw angle (θyaw), focal lengths fx, fy, optical center cx, cy, and scale factors, resulting in an 8-dof model."

    The 8-DoF count is measured from the authors' own 447 training clips (Table 1) by dropping tz, pitch, and roll because their standard deviations are small. This data-derived number is then used as the required dimension d in the theoretical proof to set the LoRA rank threshold r≥8. Thus the 'prediction' that rank ≥8 suffices is not an independent theoretical result; it is a restatement of the authors' own empirical simplification. The theory does not derive 8 from first principles; it takes it as an input. The paper's own ablation (Table 3) shows rank 8 gives L-R Cons. 0.95, below the rank-16 value of 0.99, so the predicted threshold is not even validated by the authors' experiments.

  2. self definitional [Section 3.3, Theoretical Proof, single-layer case]
    "For the transformation to span a subspace of dimension d, the rank r must satisfy r≥d. Since the transformation from a perspective to a panoramic view requires 8 degrees of freedom, we conclude that if the LoRA rank r≥8, the changeable space can be fully covered."

    The proof establishes only rank(δF) ≤ min(rank(J_F), rank(Δθ)) ≤ r, an upper bound on the rank of the actual output change for a given LoRA update. It then equates 'the transformation requires 8 degrees of freedom' with 'the changeable space has dimension 8' and concludes r≥8 covers it. This is definitional: the 'required space' is defined to be 8-dimensional from the camera model, so the conclusion that an 8-dimensional LoRA spans an 8-dimensional space is true by definition. It does not show that the specific perspective-to-panoramic weight update in a 14B diffusion transformer lies in a rank-8 subspace. An upper bound on rank cannot establish a lower bound on the space of achievable changes, so the conclusion reduces to the definition of the solution space rather than being derived from

full rationale

No load-bearing self-citations or imported uniqueness theorems are present; the derivation is the authors' own. However, the central theoretical claim is circular in the sense that the 8-DoF count is read off the authors' training clips (Table 1) and then used as the required dimension in the rank argument. Moreover, the rank inequality rank(δF) ≤ r is an upper bound and does not imply sufficiency: the paper never shows that the needed perspective-to-panoramic weight update lies in the rank-r subspace. The paper's own Table 3 contradicts the sufficiency claim, since rank 8 gives L-R consistency 0.95 (not the near-perfect closure implied by the theory) while rank 16 gives 0.99. Thus the 'theoretical analysis' does not independently predict the optimal configuration; it is post hoc and data-fitted. Score 6 reflects partial circularity: the rank threshold is fitted to the authors' data and the proof is a non-sequitur, though the experimental contributions and empirical comparisons stand on their own.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

The central argument rests on a data-derived DoF count (8), an unproven sufficiency link between LoRA rank and task dimension, and a locality assumption for non-linear layers. These are all assumptions the paper does not independently justify.

free parameters (2)
  • Task degrees of freedom d = 8
    The 11-DoF model (Eq. 2) is reduced to 8 by dropping tz, pitch, and roll because their standard deviations in Table 1 are small. This value directly sets the theoretical rank threshold.
  • LoRA rank r = 16 (default)
    Ablation shows rank 8 gives L-R 0.95 and rank 32 gives L-R 0.90; rank 16 is chosen empirically. The theory's threshold of 8 does not select the best rank.
axioms (3)
  • domain assumption The perspective-to-panoramic transformation for the target distribution is fully described by 8 DoF.
    Used to equate the task dimension with d=8 in the rank argument. Stated after Table 1 without evidence that pitch/roll/tz are negligible for the full target distribution.
  • ad hoc to paper The output change delta F satisfies rank(delta F) <= min(rank(J_F), rank(delta theta)) <= r and spans dimension d whenever r >= d.
    This is the core step of the proof; only the upper bound is justified and the sufficiency direction is assumed. The Jacobian's reachable subspace is never characterized.
  • domain assumption Non-linear layers (ReLU, Softmax, attention) behave linearly on the scale of the LoRA perturbation, so the linear rank analysis extends.
    Invoked in 'Extension to the non-linear networks' without quantifying the approximation error or showing that the relevant directions are preserved.

pith-pipeline@v1.3.0-alltime-deepseek · 12049 in / 14871 out tokens · 153037 ms · 2026-08-04T17:05:28.646266+00:00 · methodology

0 comments
read the original abstract

Generating high-quality 360{\deg} panoramic videos remains a significant challenge due to the fundamental differences between panoramic and traditional perspective-view projections. While perspective videos rely on a single viewpoint with a limited field of view, panoramic content requires rendering the full surrounding environment, making it difficult for standard video generation models to adapt. Existing solutions often introduce complex architectures or large-scale training, leading to inefficiency and suboptimal results. Motivated by the success of Low-Rank Adaptation (LoRA) in style transfer tasks, we propose treating panoramic video generation as an adaptation problem from perspective views. Through theoretical analysis, we demonstrate that LoRA can effectively model the transformation between these projections when its rank exceeds the degrees of freedom in the task. Our approach efficiently fine-tunes a pretrained video diffusion model using only approximately 1,000 videos while achieving high-quality panoramic generation. Experimental results demonstrate that our method maintains proper projection geometry and surpasses previous state-of-the-art approaches in visual quality, left-right consistency, and motion diversity.

Figures

Figures reproduced from arXiv: 2509.11092 by Eric Li, Hao-Xiang Guo, Yikai Wang, Yuqi Li, Yuyang Yin, Zeyu Dong.

Figure 1
Figure 1. Figure 1: Our goal is to generate high-quality panoramic videos using pretrained video diffusion models. We first analyze how to transform [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison across baselines. Each column shows four consecutive equirectangular frames. 360DVD: largely static; 360DVD+: limited dynamics with faint seam blur; GenEX: temporal degradation (distortions/blur); Ours: sharper details and smoother temporal transitions. View in color and zoom for detail. gresses. This discrepancy explains why GenEX’s motion magnitude is not accompanied by coherent ca… view at source ↗
Figure 3
Figure 3. Figure 3: Left–right boundary inspection over three consecutive frames. Our method achieves reliable seam closure while preserving fine [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Perspective comparison of four viewpoints concatenated [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Rank capacity threshold. Ranks < 8 underfit: seam breaks/repetitions and weak parallax (rank 5 worst; rank 8 partial recovery); rank 16 is most stable; rank 32 increases motion but in￾troduces drift/flicker. underfit and fail to maintain global panoramic continuity. (a) Rank 5 (b) Rank 8 (c) Rank 16 (d) Rank 32 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Effect of rank on left–right seam. Lower capacity (rank 5) cannot form a valid panoramic wrap, causing breaks at the ±180◦ boundary; increasing to rank 8 restores left–right alignment and a closed seam. Conclusion We propose a novel approach to panoramic video generation by reformulating it as a lightweight LoRA-based adaptation task, enabling efficient fine-tuning of pretrained video dif￾fusion models. Ou… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 12 linked inside Pith

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    B.; and Tulyakov, S

    Bahmani, S.; Skorokhodov, I.; Qian, G.; Siarohin, A.; Menapace, W.; Tagliasacchi, A.; Lindell, D. B.; and Tulyakov, S. 2025. Ac3d: Analyzing and improving 3d camera control in video diffusion transformers. In Proceedings of the Computer Vision and Pattern Recognition Conference, 22875--22889

    2025

  4. [4]

    Conrad, M.; Kablitz, D.; and Schumann, S. 2024. Learning effectiveness of immersive virtual reality in education and training: A systematic review of findings. Computers & Education: X Reality, 4: 100053

    2024

  5. [5]

    Farneb \"a ck, G. 2003. Two-frame motion estimation based on polynomial expansion. In Scandinavian conference on Image analysis, 363--370. Springer

    2003

  6. [6]

    Ferreira, S. B. L.; dos Santos, R. C.; da Silveira, D.; and Ferreira, M. G. L. 2007. Panorama. RCA--Revista de Controle e Administra c \ a o Volume III, n 2, jul./dez. 2007 Conselho Editorial , 3(2): 205

    2007

  7. [7]

    B.; and Dengel, A

    Frolov, S.; Moser, B. B.; and Dengel, A. 2025. Spotdiffusion: A fast approach for seamless panorama generation over time. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2073--2081. IEEE

    2025

  8. [8]

    Gao, S.; Yang, K.; Shi, H.; Wang, K.; and Bai, J. 2022. Review on panoramic imaging and its applications in scene understanding. IEEE Transactions on Instrumentation and Measurement, 71: 1--34

    2022

  9. [9]

    Y.; Taylor, D.; and Clarke, D

    Gledhill, D.; Tian, G. Y.; Taylor, D.; and Clarke, D. 2003. Panoramic imaging—a review. Computers & Graphics, 27(3): 435--445

    2003

  10. [10]

    J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; Chen, W.; et al

    Hu, E. J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; Chen, W.; et al. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2): 3

    2022

  11. [11]

    Jayaram, S.; Vance, J.; Gadh, R.; Jayaram, U.; and Srinivasan, H. 2001. Assessment of VR technology and its applications to engineering problems. Journal of Computing and Information Science in Engineering, 1(1): 72--83

    2001

  12. [12]

    Li, Y.; Shum, H.-Y.; Tang, C.-K.; and Szeliski, R. 2004. Stereo reconstruction from multiperspective panoramas. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(1): 45--62

    2004

  13. [13]

    N.; Tulyakov, S.; and Ren, J

    Liang, H.; Cao, J.; Goel, V.; Qian, G.; Korolev, S.; Terzopoulos, D.; Plataniotis, K. N.; Tulyakov, S.; and Ren, J. 2025. Wonderland: Navigating 3d scenes from a single image. In Proceedings of the Computer Vision and Pattern Recognition Conference, 798--810

    2025

  14. [14]

    N.; Zhao, Y.; and Wei, Y

    Liang, H.; Yin, Y.; Xu, D.; Liang, H.; Wang, Z.; Plataniotis, K. N.; Zhao, Y.; and Wei, Y. 2024. Diffusion4d: Fast spatial-temporal consistent 4d generation via video diffusion models. arXiv preprint arXiv:2405.16645

    Pith/arXiv arXiv 2024

  15. [15]

    Liu, J.; Lin, S.; Li, Y.; and Yang, M.-H. 2025. Dynamicscaler: Seamless and scalable video generation for panoramic scenes. In Proceedings of the Computer Vision and Pattern Recognition Conference, 6144--6153

    2025

  16. [16]

    Lu, T.; Shu, T.; Xiao, J.; Ye, L.; Wang, J.; Peng, C.; Wei, C.; Khashabi, D.; Chellappa, R.; Yuille, A.; et al. 2024. Genex: Generating an explorable world. arXiv preprint arXiv:2412.09624

    Pith/arXiv arXiv 2024

  17. [17]

    Ma, J.; Lu, E.; Paiss, R.; Zada, S.; Holynski, A.; Dekel, T.; Curless, B.; Rubinstein, M.; and Cole, F. 2024. VidPanos: Generative panoramic videos from casual panning videos. In SIGGRAPH Asia 2024 Conference Papers, 1--11

    2024

  18. [18]

    Mangrulkar, S.; Gugger, S.; Debut, L.; Belkada, Y.; Paul, S.; and Bossan, B. 2022. PEFT : State-of-the-art Parameter-Efficient Fine-Tuning methods. https://github.com/huggingface/peft

    2022

  19. [19]

    Miller, A. L. 1996. The Panorama, the Cinema and the Emergence of the Spectacular. Wide Angle, 18(2): 34--69

    1996

  20. [20]

    Ni, J.; Zhang, C.-B.; Zhang, Q.; and Zhang, J. 2025. What Makes for Text to 360-degree Panorama Generation with Stable Diffusion? arXiv preprint arXiv:2505.22129

    Pith/arXiv arXiv 2025

  21. [21]

    Paliwal, A.; Zhou, X.; Tsarov, A.; and Kalantari, N. K. 2024. PanoDreamer: Optimization-Based Single Image to 360 3D Scene With Diffusion. arXiv preprint arXiv:2412.04827

    arXiv 2024

  22. [22]

    Pintore, G.; Jaspe-Villanueva, A.; Hadwiger, M.; Gobbetti, E.; Schneider, J.; and Agus, M. 2023. PanoVerse: automatic generation of stereoscopic environments from single indoor panoramic images for Metaverse applications. In Proceedings of the 28th International ACM Conference on 3D Web Technology, 1--10

    2023

  23. [23]

    A.; Fromm, J.; and Wohlgenannt, I

    Radianti, J.; Majchrzak, T. A.; Fromm, J.; and Wohlgenannt, I. 2020. A systematic review of immersive virtual reality applications for higher education: Design elements, lessons learned, and research agenda. Computers & education, 147: 103778

    2020

  24. [24]

    Ren, X.; Shen, T.; Huang, J.; Ling, H.; Lu, Y.; Nimier-David, M.; M \"u ller, T.; Keller, A.; Fidler, S.; and Gao, J. 2025. Gen3c: 3d-informed world-consistent video generation with precise camera control. In Proceedings of the Computer Vision and Pattern Recognition Conference, 6121--6132

    2025

  25. [25]

    Slater, M.; and Sanchez-Vives, M. V. 2016. Enhancing our lives with immersive virtual reality. Frontiers in Robotics and AI, 3: 236866

    2016

  26. [26]

    Sun, W.; Chen, S.; Liu, F.; Chen, Z.; Duan, Y.; Zhang, J.; and Wang, Y. 2024 a . Dimensionx: Create any 3d and 4d scenes from a single image with controllable video diffusion. arXiv preprint arXiv:2411.04928

    Pith/arXiv arXiv 2024

  27. [27]

    Sun, X.; Chen, Y.; Huang, Y.; Xie, R.; Zhu, J.; Zhang, K.; Li, S.; Yang, Z.; Han, J.; Shu, X.; et al. 2024 b . Hunyuan-large: An open-source moe model with 52 billion activated parameters by tencent. arXiv preprint arXiv:2411.02265

    Pith/arXiv arXiv 2024

  28. [28]

    Tan, J.; Yang, S.; Wu, T.; He, J.; Guo, Y.; Liu, Z.; and Lin, D. 2024. Imagine360: Immersive 360 video generation from perspective anchor. arXiv preprint arXiv:2412.03552

    Pith/arXiv arXiv 2024

  29. [29]

    Team, Q. 2024. Qwen2 technical report. arXiv preprint arXiv:2407.10671

    Pith/arXiv arXiv 2024

  30. [30]

    Wan, T.; Wang, A.; Ai, B.; Wen, B.; Mao, C.; Xie, C.-W.; Chen, D.; Yu, F.; Zhao, H.; Yang, J.; Zeng, J.; Wang, J.; Zhang, J.; Zhou, J.; Wang, J.; Chen, J.; Zhu, K.; Zhao, K.; Yan, K.; Huang, L.; Feng, M.; Zhang, N.; Li, P.; Wu, P.; Chu, R.; Feng, R.; Zhang, S.; Sun, S.; Fang, T.; Wang, T.; Gui, T.; Weng, T.; Shen, T.; Lin, W.; Wang, W.; Wang, W.; Zhou, W....

    Pith/arXiv arXiv 2025

  31. [31]

    Wang, Q.; Li, W.; Mou, C.; Cheng, X.; and Zhang, J. 2024 a . 360dvd: Controllable panorama video generation with 360-degree video diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 6913--6923

    2024

  32. [32]

    Wang, X.; Zheng, Y.; Wan, Z.; and Zhang, M. 2024 b . Svd-llm: Truncation-aware singular value decomposition for large language model compression. arXiv preprint arXiv:2403.07378

    Pith/arXiv arXiv 2024

  33. [33]

    N.; Zhao, Y.; and Wei, Y

    Xing, K.; Liang, H.; Xu, D.; Yin, Y.; Plataniotis, K. N.; Zhao, Y.; and Wei, Y. 2025. TiP4GEN: Text to Immersive Panorama 4D Scene Generation. arXiv preprint arXiv:2508.12415

    Pith/arXiv arXiv 2025

  34. [34]

    Yang, Z.; Teng, J.; Zheng, W.; Ding, M.; Huang, S.; Xu, J.; Yang, Y.; Hong, W.; Zhang, X.; Feng, G.; et al. 2024. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072

    Pith/arXiv arXiv 2024

  35. [35]

    Ye, W.; Ji, C.; Chen, Z.; Gao, J.; Huang, X.; Zhang, S.-H.; Ouyang, W.; He, T.; Zhao, C.; and Zhang, G. 2024. Diffpano: Scalable and consistent text to panorama generation with spherical epipolar-aware diffusion. Advances in Neural Information Processing Systems, 37: 1304--1332

    2024

  36. [36]

    Yin, Y.; Xu, D.; Wang, Z.; Zhao, Y.; and Wei, Y. 2023. 4dgen: Grounded 4d content generation with spatial-temporal consistency. arXiv preprint arXiv:2312.17225

    Pith/arXiv arXiv 2023

  37. [37]

    Yu, H.; Wang, C.; Zhuang, P.; Menapace, W.; Siarohin, A.; Cao, J.; Jeni, L.; Tulyakov, S.; and Lee, H.-Y. 2024. 4real: Towards photorealistic 4d scene generation via video diffusion models. Advances in Neural Information Processing Systems, 37: 45256--45280

    2024

  38. [38]

    Zhang, S.; Zhang, Y.; Zheng, Q.; Ma, R.; Hua, W.; Bao, H.; Xu, W.; and Zou, C. 2024. 3d-scenedreamer: Text-driven 3d-consistent scene generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 10170--10180

    2024