Meta's Fundamental AI Research (FAIR) team has introduced new models designed for whole-body control tasks, which are intended to enhance experiences in the Metaverse, as well as for watermarking content created by AI.
Motivo Model
The first model, known as Motivo, utilizes an "algorithm that leverages an unlabeled dataset of motions" to handle various body control tasks. This includes motion tracking, reaching specific poses, and optimizing rewards, all without needing extra training or planning.
According to Meta, this model "achieves competitive performance" comparable to other specialized methods and is capable of demonstrating a "human-like" array of movements and actions. The company anticipates that this research will facilitate the emergence of "fully-embodied agents in the Metaverse," paving the way for more realistic interactive characters and contributing to the "democratization of character animation."
Video Seal Framework
In addition, Meta has developed Video Seal, which serves as a "comprehensive framework for neural video watermarking." This system allows for the addition of a watermark and an optional hidden message in videos, both of which are "imperceptible to the naked eye." This technology can effectively track the source of a video to verify whether it was created by AI.
Meta claims that these watermarks show "proven resilience against common video editing efforts like blurring or cropping, as well as compression algorithms commonly used when sharing content online."
New Tools and Models
Moreover, the company has shared code for Flow Matching, a multimodal model that can generate diverse outputs, ranging from images to videos, audio, and even 3D structures such as proteins. They also announced a new data-generation framework called Theory-of-Mind, along with new tools designed to assess image-generation models in terms of diversity modeling.
Meta's advancements in these areas reflect a significant step forward in the intersection of AI and the Metaverse, showcasing their commitment to enhancing digital interactions and content authenticity.
Source: Link
