PaliGemma

Extracts and recognizes text from images at multiple resolutions (224×224 to 896×896 pixels) using a SigLIP vision encoder that processes visual features into a token sequence, which is then decoded by the Gemma language model to produce accurate character-level transcriptions. The hybrid architecture enables the model to understand text within its visual context rather than treating OCR as isolated character recognition, improving accuracy on documents with complex layouts, handwriting, or degraded quality.

Processes natural language questions about image content by encoding the image through SigLIP's vision transformer to extract spatial and semantic features, then feeding both the visual tokens and the question text to Gemma's decoder, which generates natural language answers grounded in specific image regions. The architecture enables answering questions requiring detailed visual reasoning, object relationships, and scene understanding rather than simple image classification.

Provides Google Colab notebooks that enable interactive fine-tuning and inference without local GPU setup, leveraging Colab's free GPU resources and JAX runtime. Developers can run detection, content generation, and fine-tuning workflows directly in notebooks with minimal setup, enabling rapid prototyping and experimentation without infrastructure investment.

Identifies objects within images and generates their spatial locations by encoding the image through SigLIP to extract region-level visual features, then using Gemma to decode these features into structured text descriptions that include object categories and bounding box coordinates. The approach treats object detection as a text generation problem, enabling flexible output formats and the ability to describe objects using natural language rather than fixed class vocabularies.

Performs semantic and instance segmentation by encoding images through SigLIP's spatial feature extraction, then using Gemma to generate segmentation masks or semantic descriptions of pixel-level regions. The vision-language approach enables segmentation that understands semantic meaning of regions rather than treating segmentation as purely geometric pixel clustering, allowing the model to segment based on object categories, materials, or semantic concepts.

Generates natural language descriptions of image content by encoding images through SigLIP's vision transformer to extract comprehensive visual features, then decoding these features through Gemma's language model to produce fluent, contextually appropriate captions. The architecture enables generating captions of varying length and detail level, from short single-sentence descriptions to longer paragraph-length summaries, and can be fine-tuned to match specific caption styles or domains.

Enables adaptation of pretrained PaliGemma models to specific tasks (OCR, VQA, detection, segmentation, captioning) through supervised fine-tuning using JAX, which provides efficient gradient computation and distributed training across multiple GPUs. The fine-tuning process updates model weights on task-specific datasets, allowing the base architecture to specialize for improved accuracy on target domains while maintaining the hybrid SigLIP+Gemma architecture.

Processes images at three standardized resolutions (224×224, 448×448, 896×896 pixels) through SigLIP's vision transformer, which extracts visual features at the appropriate scale for the input resolution. This enables flexible input handling where higher resolutions capture finer details at the cost of increased computation, while lower resolutions enable faster inference with reduced memory requirements, allowing developers to optimize for latency or accuracy depending on application requirements.

Provides three model variants optimized for different deployment scenarios: PaliGemma PT (pretrained, requires fine-tuning), PaliGemma FT (research-oriented, task-specific fine-tuning), and PaliGemma mix (multi-task mixture, ready for immediate use). Each variant represents a different point on the spectrum between generality and task-specificity, enabling developers to choose based on whether they have labeled data for fine-tuning or need immediate deployment.

Processes simultaneous image and text inputs by encoding the image through SigLIP to extract visual tokens and concatenating them with text embeddings from Gemma's tokenizer, then feeding the combined sequence to Gemma's decoder. This alignment approach enables the model to understand relationships between visual content and natural language queries, enabling tasks that require reasoning about both modalities simultaneously rather than treating them independently.

Distributes PaliGemma model weights and code through Hugging Face Model Hub and Kaggle Datasets, enabling open-source access without API keys or cloud infrastructure requirements. Developers can download model weights directly, integrate them into custom inference pipelines, and deploy locally or on their own infrastructure, enabling full control over inference, fine-tuning, and deployment without vendor lock-in.