On Thursday, Google unveiled PaliGemma 2, the latest version of its PaliGemma artificial intelligence vision-language model. This new iteration enhances the functionalities of its predecessor and marks an advancement in AI technology. The Mountain View company stated that this vision-language model is capable of interpreting and interacting with various visual inputs, including images and other visual assets. PaliGemma 2 is built upon the recently released Gemma 2 small language models (SLM) from August. Notably, Google claims that the new model can assess emotions depicted in the uploaded images.
Introduction of Google PaliGemma AI Model
In a blog entry, Google elaborated on the capabilities of the PaliGemma 2 AI model. While the company has a variety of vision-language models, PaliGemma was the pioneer in the Gemma family. Unlike standard large language models (LLMs), vision models incorporate additional encoders that enable them to process and convert visual content into recognizable data formats, allowing them to “see” and comprehend the external environment effectively.
The advantages of smaller vision models include their suitability for a broad range of applications, as these models are optimized for both speed and accuracy. With PaliGemma 2 being open-sourced, developers are encouraged to leverage its features in their applications.
PaliGemma 2 is available in three different parameter sizes: 3 billion, 10 billion, and 28 billion. It also supports various resolutions of 224p, 448p, and 896p, enabling optimization of the AI model’s performance for a wide array of tasks. Google claims that this model can generate detailed, contextually appropriate captions for images, identifying not just objects but also actions, emotions, and the overall narrative of scenes.
Furthermore, Google emphasized the model’s potential applications, such as in chemical formula recognition, music score identification, spatial reasoning, and the generation of chest X-ray reports. The company has also made a related research paper available in the online pre-print journal arXiv through this link.
For those interested, developers and AI enthusiasts can access the PaliGemma 2 model and its accompanying code via Hugging Face and Kaggle platforms, available here and here. The AI model supports various frameworks, including Hugging Face Transformers, Keras, PyTorch, JAX, and Gemma.cpp.
