The article introduces Dream360, a framework developed to generate diverse, high-fidelity, and high-resolution panoramas from user-selected viewpoints. This technology considers the spherical properties of 360 images, providing a more immersive and realistic environment for virtual reality applications like virtual tourism. The framework consists of two main learning stages: Spherical-VQGAN for panorama outpainting and frequency-aware refinement for improved visual fidelity. Compared to existing methods, Dream360 offers higher outpainting flexibility and fidelity, achieving lower Frechet Inception Distance scores and better visual fidelity.
Publication date: 22 Jan 2024
Project Page: Not provided
Paper: https://arxiv.org/pdf/2401.10564
