KUAISHOU announces patent for image-to-video technology
The patent "Image-to-Video Model Generation Method, Video Generation Method and Device" applied for by KUAISHOU's affiliated company Beijing Dajia Huilian Information Technology Co., Ltd. has been published. This patent involves generating video data by inputting sample images and text descriptions, utilizing spatial modules, temporal self-attention modules, and motion adapter modules, which can control video motion information and reduce system resource consumption
According to the intellectual property information from Tianyancha, recently, KUAISHOU's affiliated company Beijing Dajia Huilian Information Technology Co., Ltd. applied for the patent "Image-Based Video Model Generation Method, Video Generation Method, and Device." The abstract shows that the video generation method described in this patent includes: inputting sample images and sample text description information into a preset image-based video model; performing spatial feature extraction processing on the sample images and sample text description information based on a spatial module to obtain the first sample feature processing result; processing the first sample feature processing result based on a temporal self-attention module to obtain the second sample feature processing result; and controlling motion information based on the first sample feature processing result using a motion adapter module to obtain the third sample feature processing result; generating predicted video data based on the second sample feature processing result and the third sample feature processing result; and updating the parameters of the spatial module and motion adapter module based on the preset video data and sample video data to obtain the image-based video model. This disclosure can control the motion information of the generated video while reducing the consumption of system resources
