PaliGemma

PaliGemma is a family of vision-language models for developers and researchers building applications that interpret images alongside text prompts. Pretrained models need fine-tuning to produce useful results; mix models can handle common tasks without further tuning.

• Image questions and captions: Ask about image content or generate captions for images and short videos.
• Object detection and text reading: Identify objects and read text embedded in images.
• Task-specific fine-tuning: Adapt a pretrained model to your image data and vision-language task, or choose FT models already tuned to specific research datasets.

The models combine SigLIP's vision model with Gemma's language model. Different model sizes and image resolutions give you choices for different workloads. You can download the weights through Kaggle and Hugging Face.

The associated big_vision research code uses JAX/Flax and supports training on GPU machines you control or Google Cloud TPU VMs. Colab examples run in the cloud and cover detection, content generation and fine-tuning. The research code uses the Apache 2.0 license; that license applies to the code, not automatically to the model weights.

big_vision focuses on Google's research projects, and its project-specific code may not receive timely updates.

PaliGemma and PaliGemma 2 are separate checkpoint generations. For example, PaliGemma 2 3B uses the Gemma terms and requires acceptance before downloading. The Apache 2.0 license of big_vision applies to the research code, not these model weights.