1.8 Generative Media and Models as a Service: Gemini, Imagen and Veo

1.8 Generative Media and Models as a Service: Gemini, Imagen and Veo

Blueprint reference: Section 1.2, "Building solutions and tuning models for specific use cases (e.g., Gemini, Imagen, Veo, and models as a service in Model Garden)."

The exam does not ask you to write prompts. It asks you to pick the right generative family for a described business outcome and then to architect around that family's operational shape. Three of those shapes are quite different from each other.

The Three First-Party Families

Gemini is the general-purpose reasoning and understanding model, and the only one of the three that natively consumes multimodal input to produce text. When a scenario says "extract fields from a scanned invoice," "answer questions about this video," or "classify this support ticket," the family is Gemini even though the input is an image or video.

Imagen produces and manipulates images. Its capabilities beyond straight text-to-image are the ones that show up in scenarios: inpainting (replace a masked region), outpainting (extend beyond the original frame), upscaling, and subject or style customization from a handful of reference images. That last one is the answer whenever the requirement is brand consistency — a specific product, mascot, or house visual style reproduced across many generated assets.

Veo generates video and behaves like a batch job. A request returns an operation handle; the client polls or receives a completion signal later. Any architecture that blocks an HTTP request on video generation is wrong.

Models as a Service

Models as a service (MaaS) means calling a partner or open model through the Agent Platform's own API surface rather than through the vendor's. The practical consequences are what the exam tests:

• One IAM model. Access is granted with Google Cloud IAM rather than a separately-managed vendor API key.
• One bill. Usage appears on the Google Cloud invoice, inside existing budgets and committed-spend agreements.
• One network path. Traffic stays within Google Cloud's boundary rather than egressing to a third-party endpoint.
• No infrastructure. Unlike deploying an open checkpoint yourself, MaaS models are managed endpoints with no accelerators to size or pay for while idle.

When a scenario complains about managing separate vendor keys, unpredictable third-party invoices, or security review of external endpoints, MaaS is the intended answer.

Tuning Across the Families

Customization means different things per family, and conflating them is a common error.

• Gemini — supervised tuning on prompt/response pairs, and adapter-based tuning that produces a tuned endpoint. Teaches behaviour, format, and domain style.
• Imagen — subject customization and style customization from reference images. Teaches what a thing looks like, not how to reason.
• Veo — customization is primarily prompt-level and conditioning-image-level; there is no general-purpose supervised tuning workflow comparable to Gemini's.
• MaaS partner models — tuning availability depends entirely on the partner; assume none unless the model card says otherwise.

Architecting Around Long-Running Generation

Media generation belongs in an asynchronous pattern. The reference shape:

Client request
   └─> API service writes a job row (status = QUEUED) and publishes to Pub/Sub
          └─> Worker (Cloud Run) calls Imagen or Veo, receives an operation handle
                 └─> Worker polls or awaits completion, writes the asset to Cloud Storage
                        └─> Job row updated to READY; client notified or polls status

Three reasons this matters on the exam. First, timeouts: a synchronous HTTP request cannot wait out a video render. Second, cost control: a queue lets you apply per-user quotas and deduplicate identical requests before spending on generation. Third, retries: generation failures are common enough that idempotent job records beat naive client retries, which otherwise pay twice for the same asset.

Provenance and Safety on Generated Media

Two controls appear in governance-flavoured questions:

• SynthID embeds an imperceptible watermark into generated images and video, allowing later verification that an asset was AI-generated. This is the answer to "how do we prove which assets our systems produced" and to internal misuse-detection requirements.
• Safety filters screen both prompts and generated output. Configurable thresholds control blocking behaviour, and requests can be rejected at the prompt stage rather than after paying for generation.

For text-based generative applications, Model Armor adds prompt-injection and jailbreak screening on top of content safety filtering; that control is covered with the rest of Section 6.

Choosing the Family: A Fast Decision Table

Exam Traps

• Using Imagen to "understand" an image. Understanding is Gemini; Imagen generates and edits.
• Synchronous video generation. Veo is a long-running operation.
• Assuming partner models can be tuned. Check the model card.
• Confusing MaaS with self-deployment. MaaS has no endpoint to size; self-deployment does.
• Forgetting per-asset billing. Image and video generation is not billed per token, so token-based cost controls do not apply.