Domain 3 · Implement computer vision solutions
Vision on this exam is not the classifier-training exercise it once was. It splits into generating visual content, understanding visual content with multimodal models, and applying responsible AI to both. At 10–15% it is one of the lighter domains, but it carries a distinctive safety topic — prompt injection through images — that appears nowhere else.
Image and video generation
The generation side covers producing images from text prompts and from reference media, and likewise videos from text or reference media. Reference media matters because it is the difference between inventing an image and producing a variation consistent with something you already have.
Editing is the other half. Image-editing workflows include inpainting, mask-based edits, and prompt-driven modifications, and generated videos can be edited as well. Inpainting and masking are worth distinguishing clearly: a mask states where the change applies, while the prompt states what the change is. Throughout, the exam expects the platform's generation and editing controls to be applied appropriately rather than exhaustively.
Multimodal understanding
Understanding runs the other direction — image in, text out. Multimodal models analyze visual context and produce captions, either concise or detailed, over a single image or several at once. Building on that, you implement question-answering grounded in visual evidence, where the answer must be traceable to what is actually in the image, and generate alt-text and extended descriptions aligned to accessibility guidelines. The accessibility framing is deliberate: alt-text has an intended audience and a standard to meet, not merely a length.
For structured extraction rather than free-form description, Azure Content Understanding in Foundry Tools extracts visual characteristics and supports video analysis workflows. Two choices recur in questions: single-task versus pro-mode Content Understanding pipelines, and identifying objects, components, or regions within images and video. Pro mode is the option to reach for when a single pass over the asset is not enough for the complexity of the extraction.
Responsible AI for multimodal content
Visual content introduces failure modes that text-only safety does not cover.
The baseline is applying filters that classify unsafe or disallowed visual content. Layered on top is indirect prompt injection — instructions embedded as text inside an image, which a multimodal model will read and may follow even though no user typed them. This is the vision-specific attack the exam expects you to recognise and mitigate.
Finally there is visual policy enforcement: watermarks, prohibited symbols, brand-usage rules, and inappropriate-content detection. These are organisational constraints rather than model-safety ones, and they apply to generated and ingested imagery alike.
Quiz · Domain 3
A developer needs to generate marketing images from text prompts using Azure OpenAI. The previously used dall-e-3 deployment is now non-functional. Which model series should the developer deploy instead?