Light-field image generation with a single image based on hybrid diffusion priors for three-dimensional light-field displays

Three-dimensional (3D) light-field displays can provide natural stereoscopic perception and intuitive visual experiences, driving widespread interest in recent years. High technical barriers and complex input requirements for 3D light-field content generation make it challenging to efficiently and conveniently generate light-field content. Here, a light-field image generation method with a single image based on hybrid diffusion priors for three-dimensional light-field displays is presented. The spatial-frequency hybrid score distillation strategy leverages 2D diffusion priors in the frequency domain to enhance appearance quality while utilizing 3D diffusion priors in the spatial domain to ensure geometric consistency. The reference view guidance loss further optimizes visual quality and geometric accuracy through joint supervision of appearance and depth. With the real-time light-field content rendering and interaction pipeline, 3D light-field images can be efficiently synthesized from extracted smooth and textured meshes. Extensive experimental results demonstrate that the proposed user-friendly method can efficiently generate 3D light-field content with high visual fidelity and geometric realism from a single input image within tens of seconds, while real-time light-field content rendering and interaction can be realized. Additionally, the method offers several extended applications to further enhance the flexibility of 3D light-field content generation.