Meta has made significant strides in artificial intelligence with the public release of several state-of-the-art models by their Fundamental AI Research (FAIR) team. The initiative aims to spur future research and innovation within the AI community, providing tools that can be applied at scale globally. The release includes five models, such as image-to-text and text-to-music generation models, a multi-token prediction model, and a technique for detecting AI-generated speech.
Through this sharing, Meta hopes to inspire further advancements and responsible AI applications. One of the models, Chameleon, is a family of mixed-modal models that can process and generate both images and text simultaneously. Unlike traditional large language models (LLMs), which are typically unimodal, Chameleon can integrate any combination of text and images, making it useful for creative tasks like generating captions for images or creating new scenes from mixed media prompts.
The FAIR team has also introduced a new multi-token prediction approach for training large language models, enhancing their efficiency. This method allows models to predict multiple future words simultaneously, accelerating training. Another model, JASCO, provides more control over AI-generated music.
Meta’s FAIR releases new AI models.
JASCO can incorporate chords and beats, offering improved control over the musical output. This model performs comparably to existing baselines in terms of quality while providing significantly enhanced versatility.
Meta has also developed AudioSeal, an audio watermarking technique for detecting AI-generated speech. AudioSeal can identify specific AI-generated segments within longer audio clips, significantly improving detection speed and efficiency compared to traditional methods. This tool is ideal for large-scale and real-time applications and is released under a commercial license to help mitigate the misuse of generative AI tools.
Meta has developed automatic indicators to evaluate potential geographical disparities to ensure fairness and representation in text-to-image models. They collected over 65,000 annotations to understand regional perceptions of geographic representation in AI-generated images. Meta is releasing the evaluation code and annotations to help the community enhance model diversity.
By making these advanced models public, Meta continues to support the AI community in generating responsible, robust, and innovative solutions. The release of these models marks an important milestone in advancing artificial intelligence and its potential applications across various domains.
