A hands-on tutorial covering how to build multimodal AI applications using Google's Gemma 4 model with Hugging Face Transformers. The guide walks through six practical tasks: screenshot-to-HTML code generation, video understanding with audio extraction, multimodal function calling (tool use), object detection with bounding box visualization, image captioning, and audio understanding. All tasks share a consistent four-step pipeline — define the multimodal message, process inputs, generate outputs, decode and parse — making it straightforward to switch between modalities. Key implementation details include using AutoModelForMultimodalLM and AutoProcessor, enabling thinking mode for reasoning-heavy tasks, deterministic decoding for structured outputs, and parsing JSON bounding box predictions from model responses.
