LLM Vision and Gemini: Home Assistant camera setup

This German-language tutorial configures LLM Vision in Home Assistant to turn camera images into written event descriptions. Daniel uses Google Gemini for the walkthrough and identifies Ollama as a local LLM option. Using a local model avoids sending camera media to an external inference provider. The demonstrated Gemini and Telegram workflow is cloud-connected. The demonstrated setup uses Google's API rather than local inference.

The installation requires HACS and the separate LLM Vision Card extension. Daniel adds the card through a custom dashboard repository, creates an API key in Google AI Studio and configures the provider. He reports an error with the prefilled Gemini 2.0 Flash model and selects Gemini 2.5 Flash instead. This is his setup experience, not a universal compatibility claim.

The walkthrough adapts system and title prompts for German descriptions, then adds preview and timeline cards with camera and event filters. A driveway automation analyzes images at night or while Daniel is away, checks the response for a relevant detection and sends a snapshot with its description to Telegram. Buttons control the driveway light, request another photo or mute alerts for 30 minutes.

His Netatmo cameras are not supplying events in this setup, so he polls every minute. He notes that activity between snapshots can go unnoticed. The initial dashboard examples are tests rather than captured real incidents.