Meta Releases New AI Models for Audio, Text, and Watermarking

Meta Chameleon is a family of models that can combine text and images as input and output any combination of text and images with a single unified architecture for both encoding and decoding. Unlike most current late-fusion models that use diffusion-based learning, Meta Chameleon uses tokenization for text and images. This unified approach makes the model easier to design, maintain, and scale. Key features include:

• Unified architecture: Combines text and images as input and output.
• Tokenization: Uses tokenization for both text and images instead of diffusion-based learning.
• Safety tuned: The released models are safety tuned and support mixed-modal inputs and text-only output for research purposes.

The possibilities are endless—imagine generating creative captions for images or using a mix of text prompts and images to create an entirely new scene.

The Multi-Token Prediction model improves LLMs by training them to predict multiple future words at once, instead of the traditional one-at-a-time approach. This method enhances model capabilities and training efficiency, allowing for faster speeds. Key benefits include:

• Improved efficiency: Predicts multiple future words at once, reducing the amount of text needed for training.
• Faster speeds: Allows for quicker model training and execution.
• Open science: The pre-trained models for code completion are released under a non-commercial/research-only license, enabling the research community to investigate the method and trained models independently.

AudioSeal is an audio watermarking technique designed specifically for the localized detection of AI-generated speech. It allows for pinpointing AI-generated segments within a longer audio snippet. Key features include:

• Localized detection: Focuses on detecting AI-generated content rather than steganography.
• Enhanced speed: Improves detection speed by up to 485 times compared to previous methods, making it suitable for large-scale and real-time applications.
• Robustness and imperceptibility: Achieves state-of-the-art performance in audio watermarking.

AudioSeal is released under a commercial license and is part of Meta's efforts to ensure the responsible use of generative AI tools.

Meta's research on geographical disparities in text-to-image generation systems aims to ensure that these models work well for everyone and reflect the geographical and cultural diversity of the world. Key efforts and findings include:

• DIG In indicators: Developed automatic indicators to evaluate potential geographical disparities in text-to-image models.
• Large-scale annotation study: Conducted a study with over 65,000 annotations and more than 20 survey responses per example to understand regional perceptions of geographic representation.
• Foregrounded concepts and background representations: Learned that people use specific components within an image when perceiving geographic representation.
• Vendi Score guidance: Introduced a method to increase the representation diversity of generated samples while maintaining or improving image quality and prompt-generation consistency.

These efforts are part of Meta FAIR's collaborative approach to improving the diversity and quality of text-to-image models.