Google DeepMind Just Released PaliGemma 2: A New Family of Open-Weight Vision Language Models (3B, 10B and 28B) recently introduced the PaliGemma 2 series, a new family of Vision-Language Models (VLMs) with parameter sizes of 3 billion (3B), 10 billion (10B), and 28 billion (28B). The models support resolutions of 224×224, 448×448, and 896×896 pixels. This release includes nine pre-trained models with different combinations of sizes and resolutions, making them versatile for a variety of use cases. Two of these models are also fine-tuned on the DOCCI dataset, which contains image-text caption pairs, and support parameter sizes of 3B and 10B at a resolution of 448×448 pixels. Since these models are open-weight, they can be easily adopted as a direct replacement or upgrade for the original PaliGemma, offering users more flexibility for transfer learning and fine-tuning….
