CreateAI联合中国科学院自动化所推出NeoVerse 4D世界模型

Core Viewpoint - CreateAI has officially launched the NeoVerse 4D world model in collaboration with the Chinese Academy of Sciences, which integrates diffusion and 4DGS core technologies, enabling efficient 4D modeling without the need for expensive multi-view data [1][2] Group 1: Technology and Innovation - NeoVerse utilizes 1 million segments of open single-view videos for training, allowing the construction of a general 4D world model in just 30 seconds [1] - The model achieves a leading performance in standard benchmark tests, breaking the traditional reliance on costly multi-view data [1] - It supports various applications including 4D reconstruction, precise roaming, multi-view generation, video editing, and 3D point tracking, promoting the widespread adoption of 4D technology across industries [1] Group 2: Methodology - The model innovatively reconstructs Gaussian primitives to achieve "sparse frame reconstruction and dense frame rendering," ensuring temporal and spatial consistency in dynamic scenes [2] - NeoVerse can simulate degraded rendering modes from new perspectives using efficiently reconstructed Gaussian primitives, eliminating the need for expensive multi-view data [2] - The training pipeline requires only single-view videos, significantly lowering the technical application threshold [1][2] Group 3: Industry Impact - CreateAI's collaboration with the Chinese Academy of Sciences represents a significant achievement in addressing industry pain points through technology [2] - The scalable training of large models relies on vast amounts of data, with the new method paving the way for efficient synthesis of 4D data from single-view videos [2] - The company has previously achieved notable success in computer vision, with research papers selected for prestigious conferences, reinforcing its academic and technical leadership [2]