Gimbal360: Differentiable Auto-Leveling for Canonicalized $360^\circ$ Panoramic Image Completion

Abstract

Diffusion models excel at 2D outpainting, but extending them to $360^\circ$ panoramic completion from unposed perspective images is challenging due to the geometric and topological mismatch between perspective projections and spherical panoramas. We present Gimbal360, a principled framework that explicitly bridges perspective observations and spherical panoramas. We introduce a Canonical Viewing Space that regularizes projective geometry and provides a consistent intermediate representation between the two domains. To anchor in-the-wild inputs to this space, we propose a Differentiable Auto-Leveling module that stabilizes feature orientation without requiring camera parameters at inference. Panoramic generation also introduces a topological challenge. Standard generative architectures assume a bounded Euclidean image plane, while Equirectangular Projection (ERP) panoramas exhibit intrinsic $S^1$ periodicity. Euclidean operations therefore break boundary continuity. We address this mismatch by enforcing topological equivariance in the latent space to preserve seamless periodic structure. To support this formulation, we introduce Horizon360, a curated large-scale dataset of gravity-aligned panoramic environments. Extensive experiments show that explicitly standardizing geometric and topological priors enables Gimbal360 to achieve state-of-the-art performance in structurally consistent $360^\circ$ scene completion.

Figures (13)

• Figure 1: We propose Gimbal360, a framework that auto-levels perspective inputs and performs geometry-aware $360^\circ$ 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.
• Figure 2: Overview of the Gimbal360 framework. Given a perspective sample from our Horizon360 dataset, the Differentiable Auto-Leveling module predicts a rigid correspondence field to warp the perspective image into a gravity-aligned, yaw-centered Canonical Viewing Space. During Topologically Equivariant Generation, a Siamese Consistency Loss between the standard and horizontally shifted latent streams forces the network to natively respect the continuous $S^1$ boundary.
• 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).
• Figure 4: Qualitative comparison in outdoor scene. Our method consistently produces gravity-aligned panoramas with straight horizon lines and coherent architectural structures, while competing methods exhibit horizon tilts, curved distortions, or boundary discontinuities.
• Figure 5: Qualitative in-the-wild results. We show diverse real-world inputs with unknown camera parameters or wide-angle shots with significant roll angles.