Sber open-sources Time Adapter, a module that teaches video neural networks to control the speed of action in a frame
Prepared by Sintelligent with the help of SI. Publication time restored from the build log. . Updated: .
Generative video is learning to "feel" time: another step toward clips indistinguishable from real footage, and in open access.
The module, under 1% the size of the main model, separates two previously merged parameters, frame rate (from 15 to 60 fps) and the pace of the action itself, so slow motion or speed-up can be produced without retraining the base model. It was trained on 40 thousand clips; together with Kandinsky 5.0 Video Lite, naturalness of motion rose by 29%, visual quality by 8% and text alignment by 19%. The code is posted on Hugging Face under the MIT license, and the work was presented at ICML.
Sources 2