MagicVideo-V2: Multi-Stage High-Aesthetic Video Generation
CoRR(2024)
摘要
The growing demand for high-fidelity video generation from textual
descriptions has catalyzed significant research in this field. In this work, we
introduce MagicVideo-V2 that integrates the text-to-image model, video motion
generator, reference image embedding module and frame interpolation module into
an end-to-end video generation pipeline. Benefiting from these architecture
designs, MagicVideo-V2 can generate an aesthetically pleasing, high-resolution
video with remarkable fidelity and smoothness. It demonstrates superior
performance over leading Text-to-Video systems such as Runway, Pika 1.0, Morph,
Moon Valley and Stable Video Diffusion model via user evaluation at large
scale.
更多查看译文
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要