Meta has introduced a new model designed to solve problems focused on human movement and a tool that inserts a watermark into videos generated by artificial intelligence to help identify its origin.
Motivo, Video Seal and CLIP 1.2 are just three of the new developments in generative artificial intelligence that Meta FAIR has presented this Thursday, with which it seeks to highlight its “recent innovations in the development of agents, robustness and security, and architectures that facilitate learning automatic”.
Motivo is a fundamental model developed for controlling the behavior of embodied virtual agents, that is, exhibiting a wide variety of more human-like behaviors.
To do this, the Meta FAIR team has turned to a data set of unlabeled movements and unsupervised reinforcement learning to train an algorithm and make it learn “human-like behaviors.”
“The key technical novelty of our algorithm is learning a representation that can be used to incorporate states, movements and rewards in the same latent space,” they explain on the Meta blog. The model also adapts its movement to changes in the environment, such as gravity, wind or direct disturbances, “despite not having been trained for them.”
Researchers are confident that this advancement can help create “more realistic” non-playable characters (NPCs) and democratize both character animation and the creation of new immersive experiences.
Meta FAIR has launched Video Seal, a watermark that is added to an AI-generated video, imperceptible to the naked eye. It is added “like a hidden message”, which allows the origin of the video to be traced, and which resists editing and the comprehension algorithms used to upload ‘online’ content.
Along with this tool, the technology company has also launched Meta Omni Seal Bench, a classification table that compares the performance of different watermarking methods. And it has announced the availability of CLIP 1.2, a basic high-performance language and vision coding model.
Added to this is work on the generative Flow Matching paradigm, which replaces classical diffusion and improves performance and efficiency while allowing easy generalization to complex data.
On the agent side, Meta FAIR has introduced Explore Theory-of-Mind, a program-guided adversarial data generation for theory-of-mind reasoning that enables data generation diverse, challenging, and scalable for both training and evaluating large language models.
Also new is the big concept model, which is inspired by the way humans can plan high-level thoughts to communicate. Likewise, work has advanced on the latent transformer of metadynamic bytes and on metamemory layers.