Standard diffusion models are designed to operate on flat, Euclidean planes. When these models are tasked with generating 360-degree panoramas—typically represented as 2D equirectangular projections—they encounter a fundamental topological mismatch. In a 360-degree environment, the far-left and far-right edges of the image are actually adjacent in 3D space. Because standard convolutional layers treat these edges as independent boundaries, the resulting images almost always suffer from visible "seams" and geometric discontinuities.
Recent research has focused on bridging this gap by making diffusion models "spherical-aware," ensuring that the generated content is cyclically continuous and geometrically consistent.
The Challenge of Boundary Continuity
The most immediate problem in panoramic generation is the vertical seam. In an equirectangular map, a person walking off the right edge of the frame should immediately reappear on the left. Standard U-Net architectures, which use zero-padding at the boundaries, break this illusion.
To solve this, researchers have introduced mechanisms to enforce circularity during the denoising process. In Diffusion360, Xuansong Xie and colleagues propose a "circular blending" strategy that applies adaptive weighting to ensure the left and right edges match during both denoising and VAE decoding. Similarly, Hai Wang introduced StitchDiffusion in Customizing 360-Degree Panoramas through Text-to-Image Diffusion Models, which modifies the latent denoising process to maintain cyclic continuity without requiring massive retraining.
Other approaches, such as 360PanT by Hai Wang, use "boundary continuity encoding" and "seamless tiling" to perform image-to-image translation on panoramas. By treating the latent space as a wrap-around canvas, these models prevent the diffusion process from denoising the edges independently, which is the primary cause of jarring vertical artifacts.
Baking Spherical Geometry into the Architecture
While circular padding fixes the seam, it doesn't address the distortion inherent in equirectangular projections. Near the poles of a sphere, pixels are stretched horizontally; a standard $3 \times 3$ convolution kernel covers a much smaller physical area at the equator than it does at the poles.
To address this, researchers are moving away from flat 2D convolutions toward spherical manifolds. In Spherical Manifold Guided Diffusion Model for Panoramic Image Generation, Mai Xu and Xiancheng Sun introduce SMGD, which performs convolutions directly on the $S^2$ manifold. This ensures that the model's receptive field is consistent across the entire sphere, regardless of projection distortion.
A similar intuition drives the Spherical-Nested Diffusion (SpND) model by Xiancheng Sun et al. in Spherical-Nested Diffusion Model for Panoramic Image Outpainting. They introduce "spherical noise" as a structural prior and use spherical deformable convolutions (SDC) to adapt the kernel shape to the local distortion of the equirectangular map. By baking the geometry into the math of the convolution itself, the model naturally learns that the "world" it is generating is a sphere, not a rectangle.
Outpainting the World from a Single View
A major application of this research is "outpainting"—taking a standard narrow-field-of-view (NFoV) photo and expanding it into a full 360-degree environment. This requires the model to not only imagine what is behind the camera but also to ensure that the new content connects perfectly with the original "anchor" image.
Jionghao Wang and the creators of PanoDiff (360-Degree Panorama Generation from Few Unregistered NFoV Images) achieve this by combining pose estimation with a latent diffusion model adapted for spherical geometry. This allows the system to take a few random snapshots and "fill in the gaps" to create a navigable environment.
However, generating a full sphere can lead to "visual cheating," where the image looks locally realistic but is geometrically impossible (e.g., a hallway that doesn't loop back correctly). To combat this, Dian Zheng and Cheng Zhang proposed PanoDecouple in Panorama Generation From NFoV Image Done Right. Their framework separates the task of geometric distortion guidance from content completion, forcing the model to respect the structural constraints of the 360-degree space before it begins painting the textures.
Scaling to High Resolution and Video
As the field matures, the focus is shifting toward higher fidelity and temporal consistency. TanDiT, introduced by Andrew Bond and colleagues in TanDiT: Tangent-Plane Diffusion Transformer for High-Quality 360° Panorama Generation, moves away from equirectangular pixels entirely. Instead, it decomposes the sphere into a grid of undistorted perspective views (tangent planes), processes them using a Diffusion Transformer (DiT), and then stitches them back together. This approach allows for the generation of 4K panoramas with significantly fewer artifacts.
Finally, researchers are extending these concepts to video. Argus, developed by Ali Farhadi, Noah Snavely, and others in Beyond the Frame: Generating 360° Panoramic Videos from Perspective Videos, applies these alignment and decoding techniques to transform standard videos into full 360° panoramic experiences, ensuring that the "wrap-around" remains consistent not just in space, but across every frame of time.