Frequently Asked Questions About Google DeepMind Generative Media App-Building Framework
21 answers covering everything from basics to advanced usage.
// Basics
What is Nano Banana 2 in Google DeepMind's model suite?
Nano Banana 2 is Google DeepMind's image generation and editing model, previously known as Imagen. It supports multiple aspect ratios, search grounding (generating images informed by live web search), and image-reference-based generation for character consistency. It is the model you call when your app needs to generate or edit images, and it works best when paired with Gemini-generated prompts due to shared training data.
What is VO3 and how is it different from VO3.1 Light?
VO3 is Google DeepMind's production-grade video generation model, and VO3.1 Light is its cheaper prototyping variant at approximately $0.05 per image equivalent. VO3.1 Light is designed for development and testing, while full VO3 delivers higher visual quality for production use. Both accept a starting image frame and a motion description prompt. The cost difference is roughly an order of magnitude, so always prototype with Light before committing to VO3.
What is the difference between Gemini consumer apps and the Developer API?
Gemini consumer apps (Gemini.com) are designed for the general public with no parameter control — users interact via a chat interface. The Developer API (accessed through AI Studio) gives developers full control over model selection, system instructions, tool toggles, structured outputs, and configuration export. Consumer apps are for end users; the Developer API is for building custom applications. Vertex AI is the third option, adding enterprise-grade infrastructure and data residency controls.
What are Effective Models (E2B, E4B) in the Gemma 4 family?
Effective 2B and 4B are Gemma 4's mobile-optimized models using a per-layer embedded architecture where embeddings are stored on flash memory and paged in as needed. The actual parameter count is ~2B/4B, but they perform closer to 5B/8B models. They are designed for phones, Raspberry Pis, and Jetson Nanos — any edge device where memory is constrained. Deploy them via AI Edge Gallery, Ollama, or LM Studio for fully offline, on-device AI.
// How To
How do I use the Get Code button in AI Studio?
Once your playground configuration produces acceptable output — correct model, tools enabled, prompt validated, files uploaded — click the 'Get Code' button in the AI Studio interface. It exports the complete configuration as runnable Python or TypeScript code, including model name, tool settings, system instructions, and file references. This exported code becomes your production starting point. Never hand-write API boilerplate before using this export.
How do I build a full-stack app with AI Studio Build?
AI Studio Build is a full-stack scaffolding feature analogous to v0.dev or Lovable. Write a detailed natural-language spec covering user flow, data persistence needs, authentication method (e.g. Google OAuth), and required API features. Add custom secrets like API keys in the settings panel, enable Firebase/Firestore for database, and connect GitHub for version control. Instruct the model to create separate files per feature and always add logging. Review file diffs carefully to catch unintended changes.
How do I generate music for my app using LIA 3?
LIA 3 generates 30-second clips or full 3-minute songs with lyrics via API. For app integration, use Gemini as a prompt factory: describe the desired mood, instrumentation, tempo, and lyrics in a Gemini chat session using structured outputs, then pass the generated prompt to LIA 3. LIA Real Time is a live variant that generates music indefinitely and responds to real-time prompt changes, functioning like an AI DJ for interactive applications.
How do I handle file uploads when prototyping in AI Studio?
Use the File Upload API, which handles file storage without requiring you to configure cloud buckets. Upload your file once and reference it by URI in subsequent Gemini prompts. When using chat mode, the uploaded asset persists across the entire conversation, so you can issue multiple instructions against the same document, image, or audio file without re-uploading. This removes storage infrastructure friction during both prototyping and early production.
How do I use Gemini Live for real-time voice interactions?
Gemini Live integrates speech-to-text, LLM understanding, and text-to-speech in one pipeline. It supports screen sharing, video feed input, and multilingual output via system instructions. Set the target language or dialect in your system instructions or within the conversation turn. Use the 'Get Code' export to replicate the Live session configuration — model name, system instructions, tool calls — in your production app. This gives you a real-time multimodal voice assistant without stitching separate services together.
// Troubleshooting
Why are my TTS outputs silent or ignoring my text?
The TTS model requires a read/tell instruction prefix — sending raw text without it causes the model to ignore the content entirely. Always prefix your input with 'Read this:' or an equivalent instruction. For multi-character narration, rewrite dialogue as a play-style transcript with inline style descriptions (e.g. 'fast-paced, British accent, excited') and pass the full transcript with a read instruction prefix. The model interprets inline style cues to differentiate voices.
Why are my generated characters inconsistent across different images?
You are likely relying on long-context memory rather than explicit reference images. The model cannot reliably infer character consistency from a long context containing many characters. Fix this by generating one dedicated reference image per character first, then passing only the specific reference images for characters appearing in each scene. Do not pass your entire character library — include only the references relevant to the current generation call.
Why is my video generation costing so much during development?
Video generation with VO can cost approximately $20 per run. Two fixes: first, use VO3.1 Light instead of full VO3 during development — it costs roughly $0.05 per image equivalent. Second, gate all expensive model calls behind explicit confirmation flags (safeguard checkboxes) in your notebooks or scripts so you never accidentally trigger costly generation. Also set service_tier='flex' for batch processing to reduce costs further at the expense of latency.
How do I debug issues when using AI Studio Build for vibe coding?
Always instruct the model to add logging from the start — error messages alone are insufficient for debugging. Require separate files for each feature to isolate regressions. Review file diffs after every iteration to catch unintended changes. When the model is fixing an error, watch which files it modifies to detect if it is changing unrelated logic. These practices make AI-assisted code generation tractable and prevent cascading issues across your codebase.
// Comparisons
How does building with DeepMind's models compare to using a generic LangChain pipeline?
The DeepMind framework leverages native multimodality — Gemini processes text, images, video, audio, and code simultaneously in a single call, eliminating the need for LangChain's chaining of separate specialist models. AI Studio's 'Get Code' replaces manual pipeline assembly. The Sprint Warning principle specifically cautions against building agent framework infrastructure (like LangChain wrappers) that the model may absorb natively. The framework also includes purpose-built models for video, music, and world simulation with no generic equivalent.
Should I use Vertex AI or AI Studio for my DeepMind app?
Start with AI Studio and the Developer API unless you have a specific reason not to. Vertex AI is for enterprises needing data residency guarantees (e.g. EU data staying in EU) and teams with existing GCP infrastructure and devops capacity. AI Studio offers maximum ease of entry: create an API key and build. Only migrate to Vertex AI when compliance, data residency, or enterprise infrastructure requirements are actual blockers — not aspirational ones.
How does Gemma 4 compare to running Gemini through the API?
Gemma 4 is open-weight (Apache 2.0) and runs locally or on your own infrastructure — no API calls, no data leaving your environment. Gemini is cloud-hosted with higher capability ceilings (especially Pro) but requires API access and has per-token costs. Use Gemma for on-device deployment, sovereign AI, offline use, or when you need full model control. Use Gemini when you need maximum quality, native tool integrations, or the full AI Studio prototyping workflow.
// Advanced
What is the Sprint Warning principle and why does it matter?
The Sprint Warning is a heuristic from Paige Bailey: if you see many developers rushing to build the same infrastructure category — vector databases, agent frameworks, fine-tunes, MCP servers — that is a strong signal the base model will absorb that capability natively within 6-12 months. Before investing engineering time in infrastructure, ask whether the capability gap will still exist. This has already happened with large context windows replacing RAG for many use cases, and with built-in tool use replacing custom agent frameworks.
What is the service_tier parameter and when should I use flex vs priority?
The service_tier parameter signals scheduling priority when calling Gemini models. Set 'flex' for batch or offline jobs where you accept minutes of latency in exchange for lower cost — it functions like a batch API. Set 'priority' (~2x the price) for live user-facing requests where reliability and low latency matter. During development, default to flex. Switch to priority only for production endpoints serving real-time users.
What is Genie 3 and can I use it to build games?
Genie 3 is a world model that generates interactive, playable environments frame-by-frame from a text description and character prompt. It is composed of Nano Banana 2, VO, and Gemini working together. It outputs raw pixel frames — there is no game engine or 3D assets involved. You can use it to prototype interactive experiences, but it is fundamentally a real-time visual simulation rather than a traditional game engine. Each frame is generated by the model pipeline.
What are structured outputs and why do I need them for chained pipelines?
Structured outputs force Gemini to return responses in a predefined schema (JSON, typed objects) rather than free-text. They are essential when chaining Gemini's output into downstream model calls — for example, generating image prompts that feed into Nano Banana 2, or character lists that determine which reference images to include. Without structured outputs, you risk unparseable responses that break your automation pipeline. Enable them as a one-liner toggle in AI Studio.
Can I use Google Search grounding with image generation?
Yes. Nano Banana 2 supports search grounding, which means it can retrieve live web information when generating images. This is useful when the model needs factual visual references it cannot resolve from training data alone — for example, generating an image of a recently released product or a current public figure in a specific context. Enable it as a toggle in AI Studio alongside your image generation prompt.