MultiViewPano: Training-Free 360° Panorama Generation via Multi-View Diffusion and Pose-Aware Stitching

Simon Coessens, Akash Malhotra, Nacera seghouani

Keywords: 360° panorama generation, multi-view diffusion models, training-free generative models, novel view synthesis, pose-aware image stitching

TL;DR: MultiViewPano is a training-free framework that uses a pretrained multi-view diffusion model and a pose-aware stitcher to generate 360° panoramas from one or more arbitrarily posed input images.

Abstract: We propose MultiViewPano, a training-free framework for generating 360° panoramas from one or more arbitrarily positioned input images with varying fields of view. Existing panorama generation methods are limited by requirements for fixed 90° field-of-view inputs, single viewpoint assumptions, or extensive task-specific fine-tuning. Our approach addresses these limitations by leveraging a pretrained multi-view diffusion model (SEVA) to synthesize overlapping novel views along strategically sampled camera trajectories, followed by a novel pose-aware stitching algorithm that exploits known camera geometry for seamless fusion. Unlike traditional feature-based stitching, our pose-aware approach directly utilizes camera poses to determine optimal seams and blending regions. Experiments on standard benchmarks demonstrate that our method achieves competitive performance with state-of-the-art approaches (FID of 12.82 on Laval Indoor, 22.99 on SUN360) while supporting arbitrary input configurations without requiring retraining or fine-tuning. \href{https://anonymous.4open.science/r/MultiViewPano-71D7/README.md}{Anonymous Repository link.}

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 4886