Meta's Fundamental AI Research (FAIR) team has released new models for whole-body control tasks for Metaverse and watermarking AI-generated content.
The first model, Motivo, has trained on an "algorithm that leverages an unlabeled dataset of motions," to perform a whole range of body control tasks, "including motion tracking, goal pose reaching, and reward optimization, without additional training or planning."
Meta says the model "achieves competitive performance" with other task-specific methods and can exhibit a "human-like" range of motions and behaviors. In the future, the company believes this research will enable "fully-embodied agents in the Metaverse," and will lead to more lifelike interactable characters and "democratization of character animation."
Video Seal is a "comprehensive framework for neural video watermarking." It can add a watermark and an optional hidden message to videos. These are "imperceptible to the naked eye," and can be used to track a video's origin to determine if it's AI-generated.
Meta says these watermarks have "proven resilience against common video editing efforts like blurring or cropping, as well as compression algorithms commonly used when sharing content online."
The company also released code for Flow Matching, a multimodal model that can generate a range of outputs from images to videos, audio, and 3D structures like proteins. It also announced a data-generation framework called Theory-of-Mind and new tools for evaluating image-generation models for diversity modeling.