Meta FAIR has announced the public release of six new AI research artifacts, aiming to advance the state of the art in AI through open research and collaboration. These releases cover a range of AI applications, from multimodal generation to audio watermarking, and reflect Meta's commitment to openness and innovation in AI development.
Among the key releases is Meta Chameleon, a family of models capable of processing and generating combinations of text and images. Components of the 7B and 34B models are being released under a research-only license. The company is also sharing pre-trained models for code completion using a new Multi-Token Prediction approach, which predicts multiple future words at once to improve efficiency and capabilities.
In the realm of creative AI, Meta is introducing JASCO (Joint Audio and Symbolic Conditioning), a text-to-music generation model that accepts various conditioning inputs for improved control over musical outputs.
To address responsible AI use, the company has developed AudioSeal, an audio watermarking technique designed to detect AI-generated speech within longer audio snippets.
Meta has also supported the release of the PRISM dataset, which maps sociodemographics and preferences of diverse participants in conversations with various LLMs. Additionally, the company has shared research on measuring and improving geographical disparities in text-to-image generation systems.
These releases align with Meta's principles of openness, collaboration, excellence, and scale in AI research. The company emphasizes responsible development and hopes to inspire further innovations in the AI community.