Vision and Image Generation

AI-901 expects you to build visual AI solutions in Microsoft Foundry, not repeat old AI-900 answers. For objective 2.3, focus on multimodal models, current image generation models, Content Understanding, and the limited cases where classic vision services still fit.

Multimodal vision Image generation Service choices Exam traps Cheat sheet

Interpret visual input with a multimodal model

When the objective says to interpret visual input in prompts by using a deployed multimodal model, it means you send an image as part of the prompt and the model answers in text. You are not calling a separate image analysis endpoint first. You are asking a deployed model to reason over the image and the written instruction together.

In Microsoft Foundry, the vision-enabled families to recognize are the GPT-5.x series, GPT-4.1, GPT-4o, and the o-series. The question will usually describe a deployed model, an image input, and a text response. The correct answer is a multimodal model when the task needs flexible visual reasoning, natural-language questions, or a mixture of text and images.

Exam wording If the prompt includes a photo, chart, screenshot, receipt image, or diagram and asks for a text answer, think multimodal model first. The image is input. The answer is text.

What you can ask	Example prompt	What the model returns
Describe an image	Describe the safety issue in this warehouse photo.	A plain-language description of visible objects, context, and likely concern.
Answer a question about a chart or diagram	What trend does this sales chart show after April?	A text answer based on labels, shapes, axes, and visual pattern.
Read text visible in a photo	What warning text appears on this sign?	The visible wording, often with surrounding interpretation.
Spot a defect	Compare this product image to the reference and identify defects.	A description of visible differences, damage, missing parts, or anomalies.
Compare two images	What changed between these two screenshots?	A comparison of layout, content, state, or visible controls.
Extract structured detail	List the item name, color, quantity, and visible damage as JSON.	A structured answer when the prompt asks for fields, though not a guaranteed service schema.

This approach now covers many tasks that used to require a dedicated vision service: captions, visual question answering, light OCR, visual comparison, and diagram explanation. It is especially strong when the user wants a conversational answer or when the image must be interpreted with business context from the prompt.

It is weaker when the app needs precise pixel-level geometry, guaranteed typed output, high-volume deterministic OCR, or strict extraction fields with confidence and grounding. For those cases, compare the task with Azure Content Understanding in Foundry Tools, Azure Document Intelligence in Foundry Tools, Azure Vision in Foundry Tools, or Azure Face.

Create images and video

Image generation on AI-901 centers on Foundry Models. The current Azure OpenAI image generation offering is the gpt-image-1 family. Foundry Models also includes FLUX models from Black Forest Labs for image generation. For video generation, recognize sora as a preview video model family.

An image prompt should say what to create and how it should look. You control the subject, setting, visual style, composition, aspect, mood, and important constraints. Good prompts are specific without pretending the model is a design tool with perfect control over every pixel.

Prompt controls

Subject: the object, person type, scene, product, or concept.
Style: photo, diagram, illustration, flat icon, realistic render, or brand-safe style guidance.
Composition: foreground, background, camera angle, lighting, and empty space for text.
Aspect: landscape, portrait, square, banner, or other layout intent.

Model choices

gpt-image-1 family for current Azure OpenAI image generation.
FLUX models from Black Forest Labs for partner image generation options.
sora for text-to-video generation in preview scenarios.

Trap DALL-E, which the old AI-900 taught, has been superseded by the gpt-image-1 family. If an AI-901 question asks for the current Foundry image generation answer, choose gpt-image-1, not DALL-E.

Generated visuals also bring responsible AI requirements. Content filters help block harmful input and output. Provenance and disclosure help users understand when an image or video was generated by AI. For business and exam scenarios, avoid generating realistic images of real people without permission, avoid impersonation, and avoid protected material such as copyrighted characters, logos, or proprietary art styles when the request would misuse them.

For the exam, do not memorize creative prompt tricks. Memorize the service boundary: Foundry Models generate images and preview video; multimodal models interpret visual input; Content Understanding extracts structured fields from visual or mixed content.

Classic vision services and their current status

This is where AI-901 catches candidates who studied AI-900. The old answer was often Azure AI Vision Image Analysis, Custom Vision, or DALL-E. AI-901 expects the current Foundry view. Classic services still exist, but their status matters.

Service	Current status	Use it when	Prefer instead
Azure Vision in Foundry Tools	Active umbrella service, but Image Analysis 4.0 is deprecated and retires September 25, 2028. Read/OCR is labeled legacy.	You maintain an existing solution or need a supported legacy capability during transition.	Use a deployed multimodal model for flexible visual Q&A, or Content Understanding for schema-driven extraction.
Azure Face	Active but Limited Access. You must apply for approval. Emotion and gender inference are retired for responsible AI reasons.	You need approved face detection, verification, identification, liveness detection, or redaction.	Use Content Understanding or multimodal vision when the task is not face recognition or liveness.
Azure AI Custom Vision	Active but retiring September 25, 2028.	You maintain an existing image classification or object detection solution during migration.	Use multimodal models for flexible recognition, or Content Understanding when you need fields from images.
Azure AI Video Indexer	Active separate service for prebuilt video insights.	You need prebuilt video insights such as faces, labels, transcripts, OCR, topics, and scene information.	Use Content Understanding when you need schema-driven field extraction from video.

Decision shortcut

If the scenario asks for	Choose	Why
Natural-language reasoning over an image	Deployed multimodal model	It accepts image input and text instructions in the same prompt.
Structured fields from documents, images, audio, or video	Content Understanding	It returns JSON that follows a schema you define, with confidence and grounding.
Legacy image analysis or OCR endpoint compatibility	Azure Vision	The service exists, but parts are deprecated or legacy.
Face verification, identification, liveness, or redaction	Azure Face	It is the purpose-built face service, but gated by Limited Access.
Prebuilt video indexing insights	Azure AI Video Indexer	It provides ready-made video metadata without designing an extraction schema.

Traps candidates bring from AI-900

Trap AI-901 is not AI-900. Do not answer vision questions as if Image Analysis, Custom Vision, and DALL-E are the default current choices.

Old answer pattern

Use Image Analysis for visual description, Custom Vision for custom recognition, and DALL-E for image generation.

AI-901 answer pattern

Use a vision-enabled multimodal model for image prompts, Content Understanding for structured extraction, gpt-image-1 for image generation, and classic services only when their current status fits.

Do not treat OCR as the whole vision story. A multimodal model can read visible text, but high-volume deterministic OCR may still call for a specialized extraction service.
Do not use Face for general image understanding. Face is for approved face scenarios, including liveness and redaction.
Do not choose emotion or gender inference. Those Face capabilities were retired for responsible AI reasons.
Do not ignore retirement dates. Image Analysis 4.0 and Custom Vision are not the future-facing default answers.
Do not use video generation for video analysis. sora creates video. Video Indexer and Content Understanding analyze video.

The night-before cheat sheet

Remember	AI-901 answer
Vision-enabled model families	GPT-5.x, GPT-4.1, GPT-4o, and o-series.
What visual prompts do	Send an image with a text instruction; receive a text answer.
Common image questions	Describe, answer chart questions, read visible text, spot defects, compare images, extract details.
Current image generation	`gpt-image-1` family; FLUX models are also in Foundry Models.
Video generation	`sora` for preview text-to-video scenarios.
Azure Vision	Exists, but Image Analysis 4.0 is deprecated and Read/OCR is legacy.
Azure Face	Active, Limited Access; face detection, verification, identification, liveness, redaction.
Custom Vision	Active but retiring September 25, 2028.
Video Indexer	Active for prebuilt video insights.
Best extraction cross-over	Use Content Understanding when the output must match a schema.

Drill: Vision prompts Drill: Image generation Next: Information extraction →