Let's go Bananas with GenMedia — Guillaume Vernade, Google DeepMind

2026-05-18 · Source: AI Engineer · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, extended

Summary

DeepMind's GenMedia models, including Nano Banana 2, VIO 3.1 Light, and LIA, offer advanced multimodal generation capabilities for images, videos, and music. Nano Banana 2 supports new aspect ratios and image grounding for enhanced search, while VIO 3.1 Light provides a cost-effective solution for video generation at $0.05 per second. LIA, the music generation model, can create 30-second clips or full 3-minute songs, with a real-time variant allowing dynamic music changes. The presentation highlighted a practical application: illustrating an open-source book using Gemini to generate prompts and GenMedia models to produce corresponding visual and auditory content. It also detailed the distinction between Google's AI Studio Gemini API and Vertex AI, emphasizing the former's developer-friendly approach and the latter's enterprise-grade control.

Key takeaway

For AI Engineers building multimodal applications, you should explore DeepMind's GenMedia models, particularly by integrating Gemini for prompt generation to ensure content consistency and quality. Consider using the Interactions API for stateful context management to optimize performance and cost, especially when working with large inputs like entire books. Be mindful of regional model availability and associated costs, opting for cheaper "light" models for iterative development before upscaling.

Key insights

DeepMind's GenMedia models offer multimodal AI for creative content generation, leveraging Gemini for intelligent prompting.

Principles

• Multimodal input and output are central to "world model" vision.
• Developer advocacy ensures real-world product utility.
• Iterative prompt refinement improves generative AI output.

Method

Utilize Gemini to generate structured prompts for characters and scenes from a book, then feed these prompts to GenMedia models (Nano Banana 2, VIO, LIA) to create consistent images, videos, and music, optionally incorporating character references for visual consistency.

In practice

• Use Gemini to generate prompts for GenMedia models.
• Employ chat mode for context-aware content generation.
• Experiment with LIA Real-time for dynamic, adaptive music.

Topics

• GenMedia Models
• Multimodal AI
• Developer Advocacy
• Gemini API
• LIA Music Generation

Best for: AI Engineer, Machine Learning Engineer, AI Student

Related on AIssential

• Gemini Omni: AI Video Generation Inside Gemini — Video generation is now an integrated capability within multimodal AI assistants, not a separate tool.
• AI Music Generation Goes Consumer with Google’s MusicFX DJ — Google's MusicFX DJ brings real-time, interactive AI music generation to consumers via the Lyria RealTime diffusion model.
• Prompt to Pipeline: Building with Google's Gen Media Stack — Paige & Guillaume, Google DeepMind — Google's generative AI stack prioritizes native multimodality, cost-efficiency, and on-device agentic capabilities across its diverse model family.
• Asset Studio is entering a new era of AI-powered creativity. — Asset Studio leverages AI, including Gemini Omni, to automate and optimize marketing asset creation from brief to A/B testing.
• Interactions API: our primary interface for Gemini models and agents — Google's Interactions API unifies Gemini model and agent interactions, enabling complex, stateful, and multimodal AI applications.
• Jump to play: Building with Gemini & MediaPipe — Combine Gemini and MediaPipe in AI Studio for rapid development of real-time, interactive, physical-world applications.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI Engineer.