Meta releases AI models for motion rendering and video watermarking – News

Meta Platforms Inc. has released two artificial intelligence models that can be used to generate motion animations and video watermarks.

The algorithms, Motivo and Video Seal, became available on Thursday. The Facebook parent also introduced two internally-developed neural network architectures. One of them, a technology called LCM, is touted as a new approach to building large language models.

Motivo, the first AI model that Meta has released, can be used to animate three-dimensional avatars of the kind often featured in virtual reality applications. The model renders avatar movements based on user-provided descriptions. It can also change the avatar’s pose: a user could, for example, instruct Motivo to have a standing avatar sit or vice versa.

The model automatically adapt animations to configuration changes. For instance, it could revise the way an avatar moves if a user adds wind to the virtual environment in which the avatar is installed.

Usually, rendering-focused AI models have to be optimized for each specific type of motion they’re used to generate. That fine-tuning requires a significant amount of resources. Meta says that Motivo doesn’t require such fine-tuning, yet provides similar output quality as algorithms optimized to render specific motions. 

The main innovation in the model is the way it ingests data. Motivo encodes information about motions and the current state of an avatar into a single latent space, a mathematical structure that AI models use to store their knowledge. The latent space also holds rewards, data points that are used to guide an AI’s training process.

Meta is one of the main players in the virtual reality headset market. The company believes Motivo could help improve the quality of VR avatars and other immersive content. “We believe this research could pave the way for fully embodied agents in the Metaverse, leading to more lifelike NPCs, democratization of character animation, and new types of immersive experiences,” the company’s researchers wrote in a blog post.

Motivo released Motivo alongside Video Seal, a machine learning tool for watermarking AI-generated videos. Watermarks created by the software are invisible to the human eye. According to Meta, they can’t be removed using common editing techniques such as blurring and cropping or by compressing a clip. 

The company previously released a similar watermarking tool for audio files. Earlier, Alphabet Inc.’s Google DeepMind lab introduced a technology called SynthID for identifying AI-generated images. Like Video Seal, SynthID generates invisible watermarks designed to be difficult to remove.

Meta released its two new AI models alongside a pair of research papers. They describe two internally-developed architectures for creating neural networks.

The first technology, Flow Matching, is designed to power AI models that generate multimedia content such as videos. It’s positioned as an alternative to the diffusion architecture that powers most video generation algorithms. Meta has already implemented Flow Matching in several of its consumer-facing generative AI tools.

“Flow Matching is a state-of-the-art generative paradigm for many modalities including generation of images, videos, audio, music, 3D structures like proteins, and more,” the company’s researchers detailed. 

Meta’s other new AI architecture is called LCP, which is short for Large Concept Model. It’s designed to power large language models. 

LLMs usually generate sentences one word fragment, or token, at a time. Models powered by Meta’s LCP architecture take a different approach. “Rather than predicting the next token, the LCM is trained to predict the next concept or high-level idea, represented by a full sentence,” Meta detailed. “Overall, the LCM outperforms or matches recent LLMs in the pure generative task of summarization.”

Image: Meta