MBE4D: Multi-View Background-Editable 4D Generation from a Single Image

Le résumé fourni par la source

Recent advances in generative modeling have extended content creation from2D images to complex 3D and 4D scenes. However, generating high-fidelity, temporally consistent 4D content from a single image remains highly challenging due to missing contextual information and unstable motion patterns. We present MBE4D, a novel framework for producing dynamic 4D content from a single input image. It employs an image matrix module to synthesize geometrically consistent and temporally coherent multi-view sequences while enabling background editing via text prompts. These image sequences, containing background, multi-view, and dynamic information, supervise the optimization of 3D Gaussian point clouds. A lightweight deformation network further extends them into 4D content. The resulting 4D representations exhibit enhanced temporal stability, improved geometric fidelity, and background consistency, effectively mitigating motion discontinuity and context loss. Extensive experiments on Objaverse-XL and Diffusion4D demonstrate that MBE4D surpasses state-of-the-art baselines in realism, efficiency, and editability, facilitating immersive AR/VR applications.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.