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∘ 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 S1 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∘ scene completion.
Figures & tables
Figure 2: Overview of Gimbal360. Perspective views are sampled from Horizon360, our gravity-aligned canonical structural prior (top left). Differentiable Projective Canonicalization (DPC, bottom left) feeds the perspective input through a geometric encoder to predict a dense correspondence field, which the rigid solver collapses into a rigid coordinate grid that warps the view into the Canonical Viewing Space. Topologically Equivariant Generation (TEG, right) denoises in this space, where circular latent encoding wraps the frozen VAE via circular-padding and cropping, and a Siamese consistency loss aligns the predictions of the base and circularly shifted streams of a shared DiT. Together these mechanisms teach the generator to respect the periodic S1 boundary (bottom right), ensuring structurally consistent and seamless panorama completions.
Indoor, 5K
Outdoor, 5K
Method
FID ↓
KID( ×102 ) ↓
O-FID ↓
FAED ↓
CS ↑
DS ↓
FID ↓
KID( ×102 ) ↓
O-FID ↓
FAED ↓
CS ↑
DS ↓
HunyuanWorld 1.0
29.23
2.1981
53.26
4.05
26.48
0.92
41.37
2.2218
69.44
7.07
29.13
1.00
WorldGen
27.43
1.9683
50.50
2.45
26.44
1.32
39.78
2.0616
67.47
6.47
29.10
1.01
DiT360
38.43
3.4848
70.61
11.30
31.09
3.76
55.17
3.7295
65.56
8.35
31.40
2.08
Ours
15.92
1.0249
30.10
1.36
30.68
0.49
23.78
0.9228
47.48
1.75
31.29
0.66
Table 1: Quantitative comparison on both indoor and outdoor scenes, with 5,000 held-out samples per domain. KID is scaled by 102 . Bold marks the best result.
Figure 3: Qualitative comparison on indoor (left) and outdoor (right) scenes. Baseline completions exhibit warped structures, mismatched distortion profiles, and boundary discontinuities, whereas ours completes structurally coherent, boundary-continuous panoramas. Please zoom in for details.
Figure 4: Qualitative results of in-the-wild generalization. Inputs include real-world captures and styled photos. Competing methods produce mismatched surroundings or warped layouts, whereas Gimbal360 completes structure- and style-consistent panoramas. Please zoom in for details.
Indoor, 0.5K
Outdoor, 0.5K
Method
FID ↓
KID ↓
O-FID ↓
DS ↓
FID ↓
KID ↓
O-FID ↓
DS ↓
w/o DPC
64.48
1.2716
93.03
1.63
69.82
1.9039
108.00
1.09
Geocalib
70.50
1.3122
100.27
1.10
78.93
2.1227
112.35
0.98
PF
72.54
1.3640
102.36
0.99
81.25
2.5321
116.56
0.94
w/o TEG
55.05
0.8106
79.49
1.75
62.74
0.8194
106.62
1.49
Ours
54.18
0.8012
79.45
0.51
63.28
0.8236
99.04
0.69
Table 2: Ablation study and anchoring baselines. Geocalib and PF denote replacing DPC with GeoCalib [ 43 ] and PerspectiveFields [ 21 ] , respectively, while keeping TEG fixed. Bold marks the best result. KID is scaled by 102 .
Figure 5: Qualitative ablation results. Removing DPC yields severe structural distortion while removing TEG breaks continuity near the periodic boundary. Our full model produces structurally coherent and boundary-continuous completions.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: 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.
Figure 7: 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 perfectly straight equator and plumb architectural verticals.
Figure 8: Out-of-Distribution Generalization.
Figure 9: Out-of-Distribution Generalization.
Figure 10: Out-of-Distribution Generalization.
Figure 11: Boundary Continuity Comparison. A zoomed-in analysis of the S1 azimuthal seam. Baseline method (left) relying on inference-time circular blending suffer from unnatural texture smoothing and relocated secondary seams at the blend edges. Gimbal360 (right) natively generates seamless boundaries with preserved high-frequency details and structural integrity.
Figure 12: Limitations and Failure Cases. (Top) Generation with humans leads to inaccurate anatomical distortions. (Bottom) An extreme telephoto input (right) provides insufficient context, occasionally resulting in completion with less coherent global geometry than a wide-angled input (left).
We present a zero-shot, training-free and optimization-free framework for generating 360 panoramic images and videos by directly injecting spherical priors into pre-trained diffusion transformers. Existing methods either rely on costly fine-tuning on scarce panoramic data that limits generalization, or leverage multi-step optimization that incurs prohibitive inference latency. We observe that contemporary generative models natively exhibit some panoramic priors from large-scale training. However, these emergent capabilities are insufficient, as the models fundamentally fail to satisfy the rigorous topological constraints imposed by equirectangular projection (ERP). We introduce a zero-shot and optimization-free approach that resolves these constraints at inference time. Spherical RoPE replaces standard rotary position embeddings: low-frequency channels are re-parameterized as 3D Cartesian coordinates to natively encode the spherical manifold, while high-frequency channels are harmonically quantized to enforce exact periodicity. Coupled with complementary Semantic Distortion classifier-free guidance (CFG) that explicitly steers geometry, we avoid retraining and inherit the full creative breadth of state-of-the-art models. Our approach generalizes across diverse backbones and 360 generation modalities. We demonstrate this across text-to-panorama using Flux.1, Flux.2, and LTX-Video backbones, achieving competitive performance against baselines, all while remaining training-free. Project page: https://orhir.github.io/SpheRoPE
Or Hirschorn, Aaron Olender, Eli Alshan +3
1Amazon Prime Video · 2Tel-Aviv University · 3Hebrew University of Jerusalem
Generating 360-degree panoramas requires both local perspective fidelity and global spherical continuity. Existing representations face a trade-off: equirectangular projection (ERP) enables seamless panorama modeling but suffers from severe distortion and a domain gap to pre-trained image generators, whereas cubemaps preserve perspective geometry but often produce inconsistent faces and visible seams. We present Squaring-the-Sphere, a panoramic image generation framework built on a novel representation CubeAtlas, which unfolds all cubemap faces into a single compact canvas. By arranging orientation-aligned, rotated polar regions alongside the equatorial faces, this representation preserves the perspective geometry of cubemaps while organizing the cross-face relationships required for seamless panorama assembly. It further enables a pretrained diffusion transformer to model the complete panorama jointly in a single pass, without altering the backbone architecture or positional encoding. While CubeAtlas has a token count comparable to ERP, its structured layout makes many dense attention interactions unnecessary. We therefore introduce a layout-aware sparse attention mask that preserves the interactions required for panorama continuity while removing geometrically irrelevant token pairs. As a result, CubeAtlas achieves better quality than ERP modeling with lower attention computation. Furthermore, we introduce a training-free super-resolution technique and a fusion scheme that enhance fine details in distortion-free perspective space before achieving the final ERP panorama. Experiments on various benchmarks demonstrate that our method improves geometric fidelity, cross-view consistency, and high-resolution visual quality over previous methods.
Panoramic imagery is increasingly used in world-generation, games, and simulation, where users may need not only photorealistic scenes but also stylized and non-photorealistic environments. Large-scale text-to-image diffusion and flow models provide broad style and semantic priors for this goal, but planar image training misaligns them with the wrap-around topology and polar regions of 360∘ panoramas represented in equirectangular projection (ERP). We present SHERPA, a lightweight adaptation framework that combines frequency-selective Circular RoPE, Circular Latent Encoding/Decoding, image-side FFN adapters, and a Dual-Path Training Scheme. Circular RoPE replaces only the seam-sensitive high-frequency horizontal RoPE band with integer-periodic harmonics while preserving the pretrained lower-frequency spectrum. The Paired Panorama Path supervises geometry, while the Unpaired Style Path uses self-supervised yaw consistency for target-free stylized prompts. As a result, SHERPA generates 360∘ panoramas across both photorealistic panorama domains and open-domain stylized prompts.