Pith. sign in

REVIEW 5 major objections 5 minor 62 references

A single unposed perspective image can be auto-leveled and completed into a gravity-aligned, seamless 360° panorama without camera parameters.

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 05:40 UTC pith:5DDWLFFH

load-bearing objection Useful framing and a plausible dataset, but the no-camera-parameters claim is a 3-DoF rotation short of what the problem needs, and the ablations don't support the story. the 5 major comments →

arxiv 2603.23179 v2 pith:5DDWLFFH submitted 2026-03-24 cs.CV

Gimbal360: Canonicalizing Planar Diffusion for Spherical Panorama Completion

Yuqin Lu , Haofeng Liu , Yang Zhou , Yihua Dai , Guiqing Li , Shengfeng He , Jun Liang This is my paper
classification cs.CV
keywords panorama completionequirectangular projectionlatent diffusionauto-levelingzero-shot camera poseS1 periodicityshift-equivariance360-degree generation
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.

The paper argues that diffusion-based panorama completion fails because ordinary perspective images and equirectangular panoramas differ in both projective geometry and topology. Gimbal360 bridges this gap by first warping any input into a canonical, gravity-aligned viewing space through a learned 3-DoF rigid projection, then training the generator to respect the circular seam via a shift-equivariance constraint. A newly curated dataset of rectified panoramas supplies the needed structural prior and ground-truth alignment flows. If the argument holds, diffusion models trained on flat images can be repurposed for structurally coherent 360° scene completion from a single unposed photo, with no camera metadata at inference.

Core claim

Gimbal360's central claim is that explicitly standardizing geometric and topological priors lets pre-trained planar diffusion models generate seamless 360° panoramas. Projective distortion is removed by predicting a dense correspondence field and then constraining it through a soft-argmin solver to a 3-DoF rigid rotation, mapping the perspective input into a Canonical Viewing Space where the horizon coincides with the equator. Topological continuity is enforced by replacing zero-padding with circular padding in the VAE and adding a Siamese shift-equivariance loss, so the network learns that circularly shifting the input shifts the output identically. The authors report that this yields state

What carries the argument

The Canonical Viewing Space — a gravity-aligned spherical coordinate system where the environmental horizon is the ERP equator — is the representation that makes planar diffusion applicable to spheres. The Differentiable Auto-Leveling module is a dense correspondence predictor followed by a soft-argmin rigid-projection bottleneck; the low-rank Jacobian of the rigid manifold filters chaotic diffusion gradients and prevents non-rigid jelly-like warps. Topologically Equivariant Generation modifies the VAE's padding to circular and trains a Siamese shift-consistency loss, forcing the diffusion transformer to be S^1-equivariant. Horizon360 supplies strictly rectified panoramas and ground-truth fl

Load-bearing premise

The method assumes that an unposed perspective photo can be brought into gravity alignment by a 3-DoF rotation alone, with no camera intrinsics; if the unknown focal length or field of view is not recoverable from the soft-argmin bottleneck, the canonicalization is geometrically incomplete.

What would settle it

Take a perspective image with a known extreme vertical field of view (e.g., 15°) and a known ground-truth panorama; if Gimbal360's output places the horizon at the wrong latitude or visibly stretches the visible content, then the auto-leveling did not recover the intrinsic scale, falsifying the no-camera-parameters claim.

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

If this is right

  • In-the-wild photos with arbitrary pitch, roll, and focal length can be completed into gravity-aligned 360° environments without user-supplied camera extrinsics.
  • Seam artifacts are resolved at training time rather than by post-hoc blending, so high-frequency detail survives across the wrap-around boundary.
  • The same canonicalization lets a model trained on rectified panoramas generalize to off-level inputs, reducing structural hallucinations like tilted horizons and curved walls.
  • The shift-equivariance objective provides a template for making any Euclidean generative backbone respect periodic structure.
  • A curated, gravity-aligned panoramic dataset enables the geometric and topological priors to be learned rather than assumed.

Where Pith is reading between the lines

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

  • If auto-leveling truly recovers the 3-DoF rotation, the same module could serve as a lightweight single-image horizon and gravity estimator, independent of the completion task.
  • A stress test worth running: feed the model a perspective image whose vertical field of view is far outside the training mixture; the 3-DoF bottleneck cannot absorb focal-length changes, so output geometry may degrade in a measurable way.
  • The shift-equivariance idea could transfer to other periodic synthesis tasks, such as 360° video or seamless texture generation, where the seam is currently patched by heuristics.
  • Because the dataset is curated to be gravity-aligned, the method's success depends on the quality of that canonicalization; residual tilt in training data would propagate directly into the learned structural prior.

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

5 major / 5 minor

Summary. Gimbal360 proposes a framework for completing 360° panoramas from a single unposed perspective image by aligning the input into a gravity-aligned Canonical Viewing Space via a Differentiable Auto-Leveling module, and by enforcing S^1 continuity through Topologically Equivariant Generation. The paper introduces a curated dataset, Horizon360, and reports state-of-the-art FID/KID/FAED/CLIPScore on indoor and outdoor benchmarks against three baselines. The central claims are that auto-leveling works without camera parameters at inference and that the shift-equivariance objective mathematically enforces seamless periodic generation.

Significance. If the results held, the paper would address a real and useful problem: adapting large planar diffusion priors to spherical panorama completion. The decomposition into canonicalization, auto-leveling, and topological equivariance is conceptually appealing, and the dataset is a potentially valuable resource. The paper includes qualitative comparisons and ablations, and the idea of constraining learned correspondence through a rigid solver is a sensible inductive bias. However, the technical formulation is underspecified, the camera-parameter-free claim is not supported by the geometry, and the empirical evidence for the topological equivariance component is internally inconsistent. The lack of equations for the core modules and the narrow comparison set prevent verification of the main contributions.

major comments (5)
  1. [Sec. 3.3 / Sec. 4] Eq. (1) defines the auto-leveling warp as z_canonical = W(z_ref, R_hat), with R_hat a 3-DoF rotation. However, a perspective image with unknown intrinsics cannot be mapped to an ERP by a rotation alone: the angular extent of the image is determined by the focal length. Section 4 explicitly samples variable vertical FOV and computes fy = 1/tan(θv/2), fx = fy/r, so the training distribution includes inputs whose intrinsic scale the 3-DoF solver cannot represent. The paper provides no equation for W, no intrinsic parameter in the solver, and no experiment that varies FOV while holding other factors fixed. Appendix D.3's admission that telephoto inputs are a failure mode is consistent with this limitation. The central claim of 'without requiring camera parameters at inference' is therefore unsupported for general inputs.
  2. [Table 1 / Sec. 5.3] The TEG ablation is not consistently favorable. On CVRG-Pano, the full model has FID 63.28 vs. 62.74 for w/o TEG, and KID 0.8236 vs. 0.8194. On Structured3D, the full model has higher KID (0.8106 vs. 0.8012). The text claims that the 'quantitative gap' demonstrates the superiority of native topological constraints, but the table shows the opposite on several metrics. This contradiction needs to be resolved, e.g., with significance testing, more metrics, or a revised interpretation.
  3. [Appendix C.2] The proof of equivariance is tautological: it posits L_shift -> 0 and then concludes that Rollδ(ϵθ(X,t)) = ϵθ(Rollδ(X),t), which is the definition of the loss being zero. Because L_shift is a soft penalty with a finite weight (λ_shift=0.5), convergence to zero is not guaranteed, and even approximate equality on the training distribution does not imply mathematical forcing. The statements in §3.4 and the abstract that TEG 'ensures' or 'enforces' S^1 continuity should be softened to 'encourages' or 'regularizes.'
  4. [Sec. 3.3] The Differentiable Auto-Leveling module is not specified to a reproducible level. No equation is given for the dense-correspondence-to-rotation soft-argmin solver, the reprojection error, the rigid spherical camera model, or the differentiable warp W. The cited [23] is a stereo-disparity method, not a rotation estimator; its adaptation to 3-DoF rotation is unclear. Without these formal definitions, the claimed 'structural gradient filter' property and the module's behavior cannot be verified.
  5. [Sec. 5.2] The comparison set is too narrow to support the state-of-the-art claim. Only HunyuanWorld, WorldGen, and DiT360 are evaluated; none is a dedicated single-image 360-degree outpainting method. The related work lists several directly relevant baselines (PanoDiffusion [49], PanoDecouple [60], CamFreeDiff [54], Matrix3D [30]) that are not included. The authors should either include these baselines or clearly justify their exclusion. As it stands, the SOTA claim is not established against the most relevant prior work.
minor comments (5)
  1. [Table 1] The header 'T able 1' contains a typo.
  2. [Appendix A.2] CFG scale of 30.0 is unusually high and likely a typo for 3.0; please verify.
  3. [Sec. 4 / Appendix B.3] The mixture-distribution parameters are inconsistent: Sec. 4 states P(Θ) = λ·N + (1-λ)·U with λ≈0.7, while Appendix B.3 gives per-parameter percentages (70/30 for pitch, 80/20 for roll, and 50/20 for FOV). Clarify the exact generative process.
  4. [Sec. 5.2 / Fig. 5] The in-the-wild generalization results are only qualitative. Reporting quantitative rotation/FOV estimation error or a user study would strengthen the robustness claim.
  5. [Sec. 5.1] The FAED metric is not defined in the text; please provide a definition or citation.

Circularity Check

1 steps flagged

One definitional tautology in the shift-equivariance proof; the empirical benchmark results and auto-leveling pipeline are not circular.

specific steps
  1. self definitional [Appendix C.2 'Proof of Equivariance' (used by Section 3.4)]
    "Assuming successful convergence (Lshift→0), the properties of the L2 norm imply: Rollδ(ϵθ(X, t)) = ϵθ(Rollδ(X), t) almost everywhere."

    The asserted equivariance equality is exactly the integrand of the L_shift loss defined in Section 3.4. The 'proof' therefore assumes the conclusion: if the shift-consistency loss reaches zero, the network is shift-equivariant by definition of that loss. This provides no independent derivation that training attains this minimum, no convergence guarantee, and no out-of-distribution bound. The 'mathematically forced' S1 continuity is a restatement of the training objective rather than a derived theorem.

full rationale

The paper's central empirical contribution—state-of-the-art panorama completion measured on held-out Structured3D and CVRG-Pano benchmarks with external metrics (FID, KID, FAED, CLIPScore)—is not circular. The Differentiable Auto-Leveling module is a supervised network trained against a ground-truth flow P_GT extracted from the canonicalized Horizon360 dataset; this is a standard learning setup, not a fitted input renamed as a prediction. The Canonical Viewing Space and Horizon360 canonicalization are design choices validated empirically, not derived from the model. The Appendix C.2 equivariance 'proof' is the one load-bearing step that reduces to its own definition: it states that if the L_shift consistency loss is zero, then the network is shift-equivariant, which is true by construction of that loss. This is a conditional tautology and does not independently establish seamless continuity, but it does not taint the benchmark comparisons. The self-citation [30] in related work is not load-bearing. The telephoto-input limitation in Appendix D.3 is a correctness caveat, not a circularity. Overall circularity is low-to-moderate, centered on the overclaimed 'proof' rather than on the empirical results.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

Central claim rests on: (i) standard LDM/DiT behavior, (ii) an unstated assumption that 3-DoF rigid rotation plus a known-or-recoverable FOV fully explains perspective-to-ERP mapping, (iii) frozen VAE can be made circular by padding override, (iv) Horizon360 canonicalization is correct, (v) soft shift-loss acts as a hard topology guarantee. No invented physical entities; the Canonical Viewing Space is a coordinate convention, not a new physical object.

free parameters (4)
  • lambda_shift = 0.5
    Weight for the Siamese shift-equivariance loss, chosen empirically (Appendix A.1). The formal proof requires L_shift -> 0, which a finite soft weight does not guarantee.
  • lambda_flow = 0.1
    Weight for the dense flow supervision loss, chosen empirically to stabilize early training.
  • pose/FOV mixture distribution parameters = lambda~0.7; pitch 70% N(0,15), 30% U(-45,45); roll 80% N(0,5), 20% U(-45,45); FOV 50% N(60,10), 20% U(45,100)
    Hand-set sampling distribution for simulated perspective crops; directly shapes the model's claimed zero-shot robustness to unconstrained inputs.
  • CFG scale = 30.0
    Inference hyperparameter, unusually high and no sensitivity analysis is provided; possibly a typo for 3.0.
axioms (5)
  • standard math Latent diffusion / DiT training behaves as described by [37] and [16] under LoRA fine-tuning.
    The L_LDM formulation and conditioning mechanism rely on standard LDM/DiT assumptions, which the paper does not re-derive.
  • domain assumption A single 3-DoF rigid rotation is a complete model of the perspective-to-ERP mapping for unposed inputs.
    Section 3.3 defines z_canonical = W(z_ref, R_hat) with no intrinsic parameter; Section 4 samples variable FOV, but no mechanism estimates FOV at inference. This is load-bearing for auto-leveling.
  • domain assumption A frozen pre-trained VAE can be made topology-preserving by overriding horizontal zero-padding with circular padding without retraining.
    Section 3.4 and A.1 state the VAE is frozen and only padding is overridden; no reconstruction-fidelity evaluation is provided, so the latent manifold may be altered.
  • domain assumption Horizon360 canonicalization via vanishing-point/zenith alignment is correct on all 20k panoramas and generalizes to web/AI images.
    Section 4/B.2 assumes the rectification pipeline always yields a gravity-aligned equator; no verification or failure analysis is reported.
  • ad hoc to paper The Siamese shift-equivariance loss with finite weight is treated as a hard guarantee of S^1 continuity.
    The Appendix C.2 proof requires L_shift = 0 exactly; the paper optimizes a soft penalty with lambda_shift=0.5, so 'mathematically forced' is an overstatement.

pith-pipeline@v1.3.0-alltime-deepseek · 15167 in / 16895 out tokens · 169046 ms · 2026-08-04T05:40:04.560673+00:00 · methodology

0 comments
read the original abstract

Diffusion models provide powerful priors for 2D image completion, but these priors are learned on bounded planar images and do not transfer directly to $360^\circ$ panoramas. Perspective observations and spherical panoramas differ in both projective geometry and topology: viewpoint-dependent distortion complicates spatial correspondence, while Equirectangular Projection (ERP) panoramas exhibit intrinsic $S^1$ periodicity that standard Euclidean architectures do not preserve. We present Gimbal360, a unified framework that adapts planar diffusion priors to spherical panoramic completion by standardizing these geometric and topological structures. Our Canonical Viewing Space expresses projective distortion as a fixed function of latitude, providing a consistent interface between perspective inputs and spherical panoramas. To map unposed in-the-wild images into this space, Differentiable Projective Canonicalization projects a dense correspondence field onto a 3-DoF rigid projection manifold without requiring camera parameters at inference. We further introduce Topologically Equivariant Generation, which enforces latent shift equivariance to preserve continuity across the periodic ERP boundary. Together, these designs allow diffusion to operate in a representation whose geometry and topology are explicitly aligned with the spherical domain. We also introduce Horizon360, a curated large-scale dataset of gravity-aligned panoramic environments. Extensive experiments show that Gimbal360 achieves state-of-the-art visual fidelity and seam continuity in $360^\circ$ scene completion.

Figures

Figures reproduced from arXiv: 2603.23179 by Guiqing Li, Haofeng Liu, Jun Liang, Shengfeng He, Yang Zhou, Yihua Dai, Yuqin Lu.

Figure 1
Figure 1. Figure 1: We propose Gimbal360, a framework that auto-levels perspective inputs and performs geometry-aware 360◦ completion in a canonical viewing space. Our method preserves global structure and maintains continuity across the periodic panorama seams (red boxes). Please zoom in for a better view. Abstract. Diffusion models excel at 2D outpainting, but extending them to 360◦ panoramic completion from unposed perspec… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the Gimbal360 framework. Given a perspective sample from our Horizon360 dataset, the Differentiable Auto-Leveling module predicts a rigid cor￾respondence field to warp the perspective image into a gravity-aligned, yaw-centered Canonical Viewing Space. During Topologically Equivariant Generation, a Siamese Con￾sistency Loss between the standard and horizontally shifted latent streams forces the … view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison in indoor scene. Challenging vertical extents stress-test geometric consistency: our method corrects tilted inputs to plumb walls and level floors, versus baseline artifacts (tilted horizons, barrel distortion). 5.2 Main Results We conduct comparison with recent state-of-the-art methods, including Hun￾yuanWorld 1.0 [18], DiT360 [10] and WorldGen [51]. Quantitative Comparison. We pres… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison in outdoor scene. Our method consistently produces gravity-aligned panoramas with straight horizon lines and coherent architectural struc￾tures, while competing methods exhibit horizon tilts, curved distortions, or boundary discontinuities. consistent perspective, stemming from its inability to handle projective variance. WorldGen, while improving upon these baselines, still suffers … view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative in-the-wild results. We show diverse real-world inputs with un￾known camera parameters or wide-angle shots with significant roll angles. 5.3 Ablation Studies We conduct extensive ablations to validate the effectiveness of two main pro￾posed mechanism, the differentiable Auto-Leveling module that anchors any perspective inputs to our Canonical Viewing Space, and Topological Equivariant Generatio… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative ablation results. Our full model achieves structurally coherent and seamless omnidirectional completion from arbitrary perspectives. 6 Conclusion We presented Gimbal360, a framework for completing 360◦ panoramas from a single unposed perspective image by explicitly addressing the geometric and topological mismatch between Euclidean diffusion priors and spherical panora￾mas. By introducing a Can… view at source ↗
Figure 7
Figure 7. Figure 7: Horizon360 Dataset Diversity. A representative sample of our curated, gravity-aligned panoramas. The dataset spans a wide domain of indoor structures and outdoor landscapes under diverse lighting conditions, providing a robust, generalized training signal for the Gimbal360 framework. and dense geometric structures, as well as expansive outdoor scenarios (e.g., natural landscapes, urban street intersections… view at source ↗
Figure 8
Figure 8. Figure 8: Dataset Canonicalization. Visualization of the rectification process applied to Horizon360. Raw unaligned panoramas (left) exhibiting severe sinusoidal distortions are mathematically rotated into the Canonical Viewing Space (right), ensuring a per￾fectly straight equator and plumb architectural verticals. B.3 Mixture-Distribution Pose Sampling To enhance zero-shot robustness when processing uncalibrated, i… view at source ↗
Figure 9
Figure 9. Figure 9: Out-of-Distribution Generalization. Gimbal360 tested on unconstrained images sourced from both real-world smartphone captures and AI-generated art. Our method robustly anchors the perspective and synthesizes coherent, high-fidelity envi￾ronments. D.2 Seamless Synthesis at S 1 Boundaries A core contribution of Gimbal360 is the mathematical elimination of topological severing. To visually validate this, [PI… view at source ↗
Figure 10
Figure 10. Figure 10: Out-of-Distribution Generalization [PITH_FULL_IMAGE:figures/full_fig_p025_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Out-of-Distribution Generalization [PITH_FULL_IMAGE:figures/full_fig_p026_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Boundary Continuity Comparison. A zoomed-in analysis of the S 1 az￾imuthal seam. Baselines relying on inference-time circular blending suffer from unnat￾ural texture smoothing and relocated secondary seams at the blend edges. Gimbal360 natively generates seamless boundaries with preserved high-frequency details and struc￾tural integrity. resolving the boundary condition mathematically within the network’s… view at source ↗
Figure 13
Figure 13. Figure 13: Limitations and Failure Cases. (Top) Generation conditioned on a per￾spective input dominated by people, leading to inaccurate anatomical distortions in the ERP format. (Bottom) Compared to standard wide-angled input (left), an extreme tele￾photo input (right) provides insufficient structural context due to small FOV, resulting in an under-constrained generation where the surrounding global geometry lacks… 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

62 extracted references · 5 linked inside Pith

  1. [1]

    In: CVPR

    Akimoto, N., Matsuo, Y., Aoki, Y.: Diverse plausible 360-degree image outpainting for efficient 3dcg background creation. In: CVPR. pp. 11441–11450 (2022) 4

    2022

  2. [2]

    In: International Conference on Learning Representations (2018) 10

    Bińkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying mmd gans. In: International Conference on Learning Representations (2018) 10

    2018

  3. [3]

    arXiv preprint arXiv:2506.21681 (2025) 4

    Çapuk, H., Bond, A., Kızıl, M.B., Göçen, E., Erdem, E., Erdem, A.: Tandit: Tangent-plane diffusion transformer for high-quality 360{\deg}panorama gen- eration. arXiv preprint arXiv:2506.21681 (2025) 4

    Pith/arXiv arXiv 2025

  4. [4]

    In: 2017 International Conference on 3D Vision (3DV)

    Chang, A., Dai, A., Funkhouser, T., Halber, M., Niebner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environ- ments. In: 2017 International Conference on 3D Vision (3DV). pp. 667–676. IEEE Computer Society (2017) 9

    2017

  5. [5]

    In: CVPR

    Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked gener- ative image transformer. In: CVPR. pp. 11315–11325 (2022) 3

    2022

  6. [6]

    ACM TOG41(6), 1–16 (2022) 4

    Chen, Z., Wang, G., Liu, Z.: Text2light: Zero-shot text-driven hdr panorama gen- eration. ACM TOG41(6), 1–16 (2022) 4

    2022

  7. [7]

    In: ECCV

    Coors, B., Condurache, A.P., Geiger, A.: Spherenet: Learning spherical represen- tations for detection and classification in omnidirectional images. In: ECCV. pp. 518–533 (2018) 4

    2018

  8. [8]

    In: ICML (2024) 3

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., Müller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al.: Scaling rectified flow transformers for high-resolution image synthesis. In: ICML (2024) 3

    2024

  9. [9]

    In: CVPR

    Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR. pp. 12873–12883 (2021) 3

    2021

  10. [10]

    arXiv preprint arXiv:2510.11712 (2025) 4, 9, 11

    Feng, H., Zhang, D., Li, X., Du, B., Qi, L.: Dit360: High-fidelity panoramic image generation via hybrid training. arXiv preprint arXiv:2510.11712 (2025) 4, 9, 11

    arXiv 2025

  11. [11]

    arXiv preprint arXiv:2311.13141 (2023) 2, 4

    Feng,M.,Liu,J.,Cui,M.,Xie,X.:Diffusion360:Seamless360degreepanoramicim- age generation based on diffusion models. arXiv preprint arXiv:2311.13141 (2023) 2, 4

    Pith/arXiv arXiv 2023

  12. [12]

    Communications of the ACM63(11), 139–144 (2020) 3

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM63(11), 139–144 (2020) 3

    2020

  13. [13]

    Guo,M.H.,Lu,C.Z.,Hou,Q.,Liu,Z.,Cheng,M.M.,Hu,S.M.:Segnext:Rethinking convolutional attention design for semantic segmentation35, 1140–1156 (2022) 7

    2022

  14. [14]

    In: Proceedings of the 2021 conference on empirical methods in natural language processing

    Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y.: Clipscore: A reference- free evaluation metric for image captioning. In: Proceedings of the 2021 conference on empirical methods in natural language processing. pp. 7514–7528 (2021) 10

    2021

  15. [15]

    Advances in neural information processing systems30(2017) 10

    Heusel,M.,Ramsauer,H.,Unterthiner,T.,Nessler,B.,Hochreiter,S.:Ganstrained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems30(2017) 10

    2017

  16. [16]

    Iclr1(2), 3 (2022) 19

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. Iclr1(2), 3 (2022) 19

    2022

  17. [17]

    arXiv preprint arXiv:2506.17206 (2025) 4

    Huang, Y., Zhou, Y., Wang, J., Huang, K., Liu, X.: Dreamcube: 3d panorama gen- eration via multi-plane synchronization. arXiv preprint arXiv:2506.17206 (2025) 4

    Pith/arXiv arXiv 2025

  18. [18]

    arXiv preprint (2025) 2, 4, 11, 23

    HunyuanWorld, T.: Hunyuanworld 1.0: Generating immersive, explorable, and in- teractive 3d worlds from words or pixels. arXiv preprint (2025) 2, 4, 11, 23

    2025

  19. [19]

    Jaderberg, M., Simonyan, K., Zisserman, A., et al.: Spatial transformer networks 28(2015) 2, 7 16 Lu et al

    2015

  20. [20]

    In: CVPR

    Jiang, Z., Xiang, Z., Xu, J., Zhao, M.: Lgt-net: Indoor panoramic room layout estimation with geometry-aware transformer network. In: CVPR. pp. 1654–1663 (2022) 9

    2022

  21. [21]

    In: CVPR

    Jin, L., Zhang, J., Hold-Geoffroy, Y., Wang, O., Blackburn-Matzen, K., Sticha, M., Fouhey, D.F.: Perspective fields for single image camera calibration. In: CVPR. pp. 17307–17316 (2023) 10

    2023

  22. [22]

    In: ICLR (2025) 4

    Kalischek, N., Oechsle, M., Manhardt, F., Henzler, P., Schindler, K., Tombari, F.: Cubediff: Repurposing diffusion-based image models for panorama generation. In: ICLR (2025) 4

    2025

  23. [23]

    In: ICCV

    Kendall, A., Martirosyan, H., Dasgupta, S., Henry, P., Kennedy, R., Bachrach, A., Bry, A.: End-to-end learning of geometry and context for deep stereo regression. In: ICCV. pp. 66–75 (2017) 7

    2017

  24. [24]

    In: ICCV

    Kocabas, M., Huang, C.H.P., Tesch, J., Müller, L., Hilliges, O., Black, M.J.: Spec: Seeing people in the wild with an estimated camera. In: ICCV. pp. 11035–11045 (2021) 9

    2021

  25. [25]

    Labs, B.F.: Flux.https://github.com/black-forest-labs/flux(2024) 10, 19

    2024

  26. [26]

    Lee, Y., Kim, K., Kim, H., Sung, M.: Syncdiffusion: Coherent montage via syn- chronized joint diffusions36, 50648–50660 (2023) 2, 4

    2023

  27. [27]

    Li, J., Bansal, M.: Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation36, 21878–21894 (2023) 4

    2023

  28. [28]

    IEEE TIP33, 382–394 (2023) 4

    Liao, K., Xu, X., Lin, C., Ren, W., Wei, Y., Zhao, Y.: Cylin-painting: Seamless 360 panoramic image outpainting and beyond. IEEE TIP33, 382–394 (2023) 4

    2023

  29. [29]

    In: ECCV

    Liu, A., Li, Z., Chen, Z., Li, N., Xu, Y., Plummer, B.A.: Panofree: Tuning-free holistic multi-view image generation with cross-view self-guidance. In: ECCV. pp. 146–164. Springer (2024) 4

    2024

  30. [30]

    In: CVPR

    Lu, Y., Zhang, J., Fang, T., Nahmias, J.D., Tsin, Y., Quan, L., Cao, X., Yao, Y., Li, S.: Matrix3d: Large photogrammetry model all-in-one. In: CVPR. pp. 11250–11263 (2025) 2, 4

    2025

  31. [31]

    In: CVPR

    Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Re- paint: Inpainting using denoising diffusion probabilistic models. In: CVPR. pp. 11461–11471 (2022) 2, 3

    2022

  32. [32]

    In: ECCV

    Oh, C., Cho, W., Chae, Y., Park, D., Wang, L., Yoon, K.J.: Bips: Bi-modal indoor panorama synthesis via residual depth-aided adversarial learning. In: ECCV. pp. 352–371. Springer (2022) 4

    2022

  33. [33]

    Signal, Image and Video Processing (2021) 10

    Orhan, S., Bastanlar, Y.: Semantic segmentation of outdoor panoramic images. Signal, Image and Video Processing (2021) 10

    2021

  34. [34]

    In: ICCV

    Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: ICCV. pp. 4195–4205 (2023) 3

    2023

  35. [35]

    Poly Haven: Poly haven hdris.https://polyhaven.com/hdris, accessed: Decem- ber 2025 9

    2025

  36. [36]

    In: ICML

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748–8763. PmLR (2021) 6

    2021

  37. [37]

    In: CVPR

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684–10695 (2022) 2, 3, 5

    2022

  38. [38]

    In: CVPR

    Sun, C., Hsiao, C.W., Sun, M., Chen, H.T.: Horizonnet: Learning room layout with 1d representation and pano stretch data augmentation. In: CVPR. pp. 1047–1056 (2019) 9

    2019

  39. [39]

    In: CVPR

    Sun, C., Sun, M., Chen, H.T.: Hohonet: 360 indoor holistic understanding with latent horizontal features. In: CVPR. pp. 2573–2582 (2021) 9 Gimbal360 17

    2021

  40. [40]

    In: ICML (2025) 4

    Sun, X., Ma, S., Li, S., Xu, M., Xia, J., Jiang, L., Deng, X., Wang, J.: Spherical- nested diffusion model for panoramic image outpainting. In: ICML (2025) 4

    2025

  41. [41]

    In: CVPR

    Sun, X., Xu, M., Li, S., Ma, S., Deng, X., Jiang, L., Shen, G.: Spherical manifold guided diffusion model for panoramic image generation. In: CVPR. pp. 5824–5834 (2025) 4

    2025

  42. [42]

    arXiv preprint arXiv:2412.03552 (2024) 4

    Tan, J., Yang, S., Wu, T., He, J., Guo, Y., Liu, Z., Lin, D.: Imagine360: Immersive 360 video generation from perspective anchor. arXiv preprint arXiv:2412.03552 (2024) 4

    Pith/arXiv arXiv 2024

  43. [43]

    In: ICCV

    Tirado-Garín, J., Civera, J.: Anycalib: On-manifold learning for model-agnostic single-view camera calibration. In: ICCV. pp. 8044–8055 (2025) 10

    2025

  44. [44]

    In: ECCV

    Veicht, A., Sarlin, P.E., Lindenberger, P., Pollefeys, M.: Geocalib: Learning single- image calibration with geometric optimization. In: ECCV. pp. 1–20. Springer (2024) 10

    2024

  45. [45]

    In: CVPR

    Wang, Q., Li, W., Mou, C., Cheng, X., Zhang, J.: 360dvd: Controllable panorama video generation with 360-degree video diffusion model. In: CVPR. pp. 6913–6923 (2024) 4

    2024

  46. [46]

    In: CVPR

    Wang, Y., Tao, X., Shen, X., Jia, J.: Wide-context semantic image extrapolation. In: CVPR. pp. 1399–1408 (2019) 3

    2019

  47. [47]

    World Labs: Generating worlds.https://www.worldlabs.ai/blog/generating- worlds, accessed: February 2026 23

    2026

  48. [48]

    In: AAAI

    Wu,T.,Li,X.,Qi,Z.,Hu,D.,Wang,X.,Shan,Y.,Li,X.:Spherediffusion:Spherical geometry-aware distortion resilient diffusion model. In: AAAI. vol. 38, pp. 6126– 6134 (2024) 4

    2024

  49. [49]

    In: ICLR (2024) 2, 4

    Wu, T., Zheng, C., Cham, T.J.: Panodiffusion: 360-degree panorama outpainting via diffusion. In: ICLR (2024) 2, 4

    2024

  50. [50]

    In: CVPR

    Xiao, J., Ehinger, K.A., Oliva, A., Torralba, A.: Recognizing scene viewpoint using panoramic place representation. In: CVPR. pp. 2695–2702. IEEE (2012) 9

    2012

  51. [51]

    Xie, Z.: Worldgen: Generate any 3d scene in seconds.https://github.com/ ZiYang-xie/WorldGen(2025) 11, 23, 24

    2025

  52. [52]

    In: Proceedings of the special interest group on computer graphics and interactive techniques conference conference papers

    Yang, S., Tan, J., Zhang, M., Wu, T., Wetzstein, G., Liu, Z., Lin, D.: Layerpano3d: Layered 3d panorama for hyper-immersive scene generation. In: Proceedings of the special interest group on computer graphics and interactive techniques conference conference papers. pp. 1–10 (2025) 9

    2025

  53. [53]

    In: ECCV

    Yao, K., Gao, P., Yang, X., Sun, J., Zhang, R., Huang, K.: Outpainting by queries. In: ECCV. pp. 153–169. Springer (2022) 3

    2022

  54. [54]

    In: CVPR

    Yuan, X., Tang, S., Li, K., Wang, P.: Camfreediff: camera-free image to panorama generation with diffusion model. In: CVPR. pp. 16408–16417 (2025) 4

    2025

  55. [55]

    In: CVPR

    Zhang, C., Wu, Q., Gambardella, C.C., Huang, X., Phung, D., Ouyang, W., Cai, J.: Taming stable diffusion for text to 360 panorama image generation. In: CVPR. pp. 6347–6357 (2024) 4, 10

    2024

  56. [56]

    In: ICCV

    Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV. pp. 3836–3847 (2023) 3

    2023

  57. [57]

    In: ECCV

    Zhang,Y.,Song,S.,Tan,P.,Xiao,J.:Panocontext:Awhole-room3dcontextmodel for panoramic scene understanding. In: ECCV. pp. 668–686. Springer (2014) 9

    2014

  58. [58]

    Zhao, S., Cui, J., Sheng, Y., Dong, Y., Liang, X., Chang, E.I., Xu, Y.: Large scale imagecompletionviaco-modulatedgenerativeadversarialnetworks.arXivpreprint arXiv:2103.10428 (2021) 3

    Pith/arXiv arXiv 2021

  59. [59]

    In: CVPR

    Zheng, C., Cham, T.J., Cai, J.: Pluralistic image completion. In: CVPR. pp. 1438– 1447 (2019) 2, 3

    2019

  60. [60]

    In: CVPR

    Zheng, D., Zhang, C., Wu, X.M., Li, C., Lv, C., Hu, J.F., Zheng, W.S.: Panorama generation from nfov image done right. In: CVPR. pp. 21610–21619 (2025) 2, 4 18 Lu et al

    2025

  61. [61]

    In: ECCV (2020) 10

    Zheng, J., Zhang, J., Li, J., Tang, R., Gao, S., Zhou, Z.: Structured3d: A large photo-realistic dataset for structured 3d modeling. In: ECCV (2020) 10

    2020

  62. [62]

    In: CVPR

    Zou, C., Colburn, A., Shan, Q., Hoiem, D.: Layoutnet: Reconstructing the 3d room layout from a single rgb image. In: CVPR. pp. 2051–2059 (2018) 9 Gimbal360 19 A Extended Experimental Setups A.1 Training Details To train Gimbal360 effectively while managing the memory footprint of the mas- sive Flux.1-fill-dev backbone [25], we employ Low-Rank Adaptation (...

    2051