ShieldGemma 2

ShieldGemma 2, released March 2025 alongside Gemma 3, is Google's safety classifier for images: a 4-billion-parameter model that scores generated or uploaded images across three harm categories — dangerous content, sexually explicit content, and violence — so applications can filter or flag them before they reach users.

The weights are open and customizable: developers can fine-tune the classifier to their own content policies and industry requirements, and it is designed to drop into generation pipelines (Imagen, Gemini image output, or any image flow) as a final safety gate. It is available on Hugging Face, Kaggle, and Google AI Studio under the Gemma license terms.

Honest caveats: it covers three image categories only — no text moderation, no fine-grained policy reasoning (that's what the larger guard models do). The Gemma license is Google's custom terms, not OSI-open. As a lightweight, tunable image-safety layer for builders already in the Gemma ecosystem, it fills a real gap.