OmniView: An All-Seeing Diffusion Model for 3D and 4D View Synthesis

📝 Original Info

• Title: OmniView: An All-Seeing Diffusion Model for 3D and 4D View Synthesis
• ArXiv ID: 2512.10940
• Date: 2025-12-11
• Authors: Xiang Fan, Sharath Girish, Vivek Ramanujan, Chaoyang Wang, Ashkan Mirzaei, Petr Sushko, Aliaksandr Siarohin, Sergey Tulyakov, Ranjay Krishna

📝 Abstract

Project page: https://snap-research.github.io/OmniView 3 views / 49 frames → 49 frames @ same time 3 views / 1 frame → 1 frame @ same time 1 view / 1 frame → 49 frames @ future time text & cam only → 49 frames @ future time 1 view / 49 frames → 49 frames @ future time Figure 1. OMNIVIEW: Given one or more images or videos, captured at potentially different viewpoints and times, OMNIVIEW generates 4D-consistent videos that can be rendered from novel camera trajectories, viewpoints, times, and durations. OMNIVIEW unifies many existing video generation paradigms, enabling tasks such as novel view synthesis (NVS), text-to-video (T2V) camera control, and multiview conditioning in one model.

📄 Full Content

...(본문 내용이 길어 생략되었습니다. 사이트에서 전문을 확인해 주세요.)