Google has unveiled PaliGemma 2, an advanced version of its vision-language model (VLM) announced earlier in 2024.
Building on the capabilities of the original PaliGemma, which focused on tasks like image captioning, object detection, and visual question answering, PaliGemma 2 introduces new features like long captioning.
This allows the model to generate detailed, context-aware captions that go beyond simple object identification, describing actions, emotions, and the overall scene.
The model also boasts improvements in optical character recognition, document table structure comprehension, and excels in tasks such as chemical formula recognition, music score interpretation, spatial reasoning, and chest X-ray report generation.
Available in multiple sizes (3B, 10B, 28B parameters) and resolutions (224px, 448px, 896px), PaliGemma 2 is designed to provide developers with an easy upgrade from the original model, offering immediate performance improvements with minimal code changes.
The pre-trained models and code are available today on platforms like Kaggle, Hugging Face, and Ollama.
