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.
Figures & tables
Figure 1: Squaring-the-Sphere reconciles global panorama continuity with perspective-face geometry. Upper: ERP represents the full panorama on a single canvas but severely stretches the polar regions; cubemaps preserve perspective geometry but introduce disconnected face boundaries. CubeAtlas packs the four equatorial faces and rotated half-polar regions into a compact canvas, making their geometric adjacencies local for joint generation. Lower: Squaring-the-Sphere produces seamless, geometrically coherent panoramas across diverse indoor and outdoor scenes, shown as ERP renderings with perspective crops.
Figure 2: CubeAtlas representation. Four equatorial faces are unfolded into the central band. The upper and lower faces are rotated into orientation-aligned half-polar copies and placed next to their corresponding equatorial faces, forming a compact layout for joint generation.
Figure 3: Overview of Squaring-the-Sphere . Given a perspective image and text prompts, the DiT denoises the complete CubeAtlas in a single pass. Layout-aware sparse attention retains within-band interactions while restricting cross-latitude communication: for each pole, an equatorial region attends only to the two axis-aligned half-polar regions that reconstruct a complete polar face. The generated CubeAtlas is then fused and stitched into the final ERP panorama.
Figure 4: The direction-aware polar blending mask.
Method
FID ↓
KID ↓
CLIP-FID ↓
Seam ERP↓
Seam Cube↓
Eq-IQ ↑
Eq-Aes ↑
Eq-GI ↑
Pole-IQ ↑
Pole-GI ↑
TextAdh ↑
Self-Collected
CubeDiff
-
-
-
0.0126
0.0361
3.685
3.565
3.465
3.780
3.725
4.100
DreamCube
-
-
-
0.0279
0.0445
2.730
2.605
2.607
2.870
2.935
3.490
HunyuanPano
-
-
-
0.0154
-
4.475
4.247
4.468
4.495
4.120
4.300
HY-Pano-2.0
-
-
-
0.0199
-
4.580
4.355
4.480
4.705
4.685
4.470
Ours
-
-
-
0.0142
0.0132
4.690
4.402
4.782
4.675
4.730
4.640
Table 1: Quantitative image-to-panorama comparison on our self-collected set, RealSee3D, Laval-Indoor, and SUN360. FID, KID, and CLIP-FID are evaluated by 90∘ perspective views split along 12 icosahedral directions (eight valid directions for Laval-Indoor) and are not reported on the self-collected set without ground-truth ERPs. Seam Cube is reported only for methods with native cube-face outputs.
Figure 5: Qualitative comparison with perspective cubemaps. Our results enjoy superior cross-face structures and pole geometry compared to ERP-native and cubemap-based baselines.
Computational Profile
Quality and Continuity
Variant
Tokens
Time
SeamERP↓
SeamCube↓
Eq-IQ ↑
Eq-Aes ↑
Eq-GI ↑
Pole-IQ ↑
Pole-GI ↑
TextAdh ↑
ERP
1.00×
1.00×
0.0209
–
4.607
4.370
4.575
4.690
4.660
4.550
Cubemap
0.75×
0.76×
0.0147
0.0260
4.428
4.270
4.332
4.490
4.455
4.560
CubeAtlas
1.00×
1.00×
0.0103
0.0143
4.697
4.353
4.690
4.705
4.705
4.640
+ Sparse Attn.
1.00×
0.84×
0.0142
0.0132
4.690
4.402
4.782
4.675
4.730
4.640
Table 2: Ablation on panorama representation and efficiency-oriented designs. “Tokens” denotes the target-token sequence length, and “Time” denotes denoising time under the same settings. Both are reported as relative multipliers normalized to the ERP baseline ( 1.00× ). “Sparse Attn.” indicates the layout-aware sparse attention used by Squaring-the-Sphere .
Figure 6: Visual comparison of panorama representations. CubeAtlas reduces cubemap face discontinuities while preserving polar structures more faithfully than direct ERP generation.
Figure 7: Qualitative comparison of panoramic super-resolution. Our method recovers sharper details while maintaining cross-face consistency.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: User study. Each stacked bar summarizes 700 valid judgments comparing Squaring-the-Sphere (solid blue) with one competitor (light fill with a dashed border). The vertical dashed line denotes the 50% equal-preference reference.
Figure 9: Examples of VLM-based evaluation. Regional examples show rationales and scores for the Front and Zenith views. “Distortion” denotes geometric integrity (GI), with higher scores indicating fewer artifacts. Equatorial views are rearranged for display, and the input caption is excerpted.
Metric
Eq-IQ
Eq-Aes
Pole-IQ
Eq-GI
Pole-GI
TextAdh
Spearman ρ
0.9358
0.8820
0.8757
0.9281
0.7698
0.8804
Appendix
Table 3: Agreement between VLM scores and human judgments. Spearman rank correlations between VLM scores and human MOS on 30 generated panoramas, each evaluated by three raters. Higher values indicate stronger rank agreement with human judgments.
Computational Profile
Quality and Continuity
Variant
Tokens
Time
Seam ERP↓
Seam Cube↓
Eq-IQ ↑
Eq-Aes ↑
Eq-GI ↑
Pole-IQ ↑
Pole-GI ↑
TextAdh ↑
Half-polar
1.00×
1.00×
0.0103
0.0143
4.697
4.353
4.690
4.705
4.705
4.640
+ Sparse Attn.
1.00×
0.84×
0.0142
0.0132
4.690
4.402
4.782
4.675
4.730
4.640
Full-polar
1.50×
1.63×
0.0106
0.0134
4.610
4.410
4.670
4.660
4.710
4.580
+ Sparse Attn.
1.50×
1.15×
0.0103
0.0136
4.620
4.410
4.720
4.660
4.680
4.620
Appendix
Table 4: Full-polar vs half-polar packing. Padding tokens are omitted from the token length counting for simplification. “+ Sparse Attn.” denotes layout-aware sparse attention.
Figure 10: Additional qualitative comparisons on Laval-Indoor, SUN360, and RealSee3D. Results are shown for Squaring-the-Sphere , HY-Pano-2.0, HunyuanPano, DreamCube, and CubeDiff.
Figure 11: Additional qualitative results for high-resolution synthesis. From left to right: Native Sampling, ERP-domain SR, perspective-space SR without overlap fusion, and the complete super-resolution pipeline. Enlarged crops compare local detail recovery and cross-view consistency.
Figure 12: Panorama generation from diverse input viewing directions. Each row shows a perspective input (left) and the generated ERP panorama (right). Our method accommodates downward- and upward-looking inputs without camera calibration or explicit input projection. Red outlines indicate the corresponding input regions in the generated panoramas.
Figure 13: More qualitative results from Squaring-the-Sphere . We show diverse indoor and outdoor scenes with realistic and stylized appearances. The generated panoramas preserve the conditioning content and visual style while maintaining coherent structures throughout the complete 360∘ FoV.
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
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.
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.