The Learning and Vision Lab team at the National University of Singapore has recently developed a distributed long-video generation tool called Video-Infinity. This tool can rapidly generate long videos using multiple GPUs without the need for additional training of existing models, producing 2300 frames of video every 5 minutes, which equates to 95 seconds in length.
Product Entry: https://top.aibase.com/tool/video-infinity
This tool features distributed, high-speed, and training-free characteristics, and is installed using a conda environment, making it user-friendly. Users can configure basic settings, pipeline settings, and Video-Infinity settings as needed to meet different video generation requirements.
Additionally, to avoid high-frequency information loss, it is recommended to set the sum of local context frames and the global context frames of the Attention model below 24 to maintain video quality and stability.
Comparative experiments have shown that this technology not only has a significant advantage in the number of video frames but also performs excellently in terms of time cost. The technical team stated that their method achieved this feat with the support of 8 Nvidia Ada6000 GPUs, and the sampling steps were set at 30.
The researchers of this technology also conducted adversarial experiments, analyzing the effects of Clip Parallelism and Dual-scope Attention through ablation studies. The results showed that these two factors have a significant impact on video generation. Furthermore, the technology supports multiple prompts (Multi-Prompts), enabling smooth transitions in background, style, and theme, providing more possibilities for video generation.
Key Points:
⭐️ This technology can generate a 2300-frame video within 5 minutes, 100 times faster than previous methods;
⭐️ Clip Parallelism and Dual-scope Attention have a significant impact on video generation;
⭐️ Supports multiple prompts (Multi-Prompts), allowing for smooth transitions.
