CertPrep Coach

AI-901 Vision and image generation practice questions

12 sample questions with full explanations. Try each one before you open the answer. New to this area? Read the Vision and image generation study guide first.

  1. Question 1

    Your app uses the Responses API with a vision-enabled model. A user uploads a photo and asks "What's damaged in this picture?". How should the app send the photo and the question?

    • A One user message with an input_text part and an input_image part
    • B Put the photo's local file path in the system message and the question in the user message
    • C Send the question in one request, then send the photo in a follow-up request
    • D Upload the photo to the model catalog first, then refer to it by name in the question
    Show answer and explanation

    Answer: A

    A single request can include text input and image input together. The image goes in an input_image content part (as a URL or base64-encoded data), next to an input_text part, and the model processes both at once.

    • B: A file path on the user's device means nothing to the model. The image itself (or a reachable URL) has to be sent.
    • C: Each request is separate, so the model answering the first request would never see the photo.
    • D: The model catalog lists models you can deploy. It doesn't store your images.

    Source: Multimodal models for image analysis (Get started with computer vision in Azure) (Microsoft Learn)

  2. Question 2

    You send images to a vision-enabled model deployed in Microsoft Foundry. For each statement, choose Yes if it is true or No if it is false.

    • An image can be provided as a URL.
    • An image can be provided as base64-encoded image data.
    • A request can contain an image or text, but not both.
    Show answer and explanation
    An image can be provided as a URL.Yes
    An image can be provided as base64-encoded image data.Yes
    A request can contain an image or text, but not both.No

    Images can be sent as URLs or as base64-encoded data, and one request can combine text and images. The model processes both inputs together to produce its response.

    Source: Multimodal models for image analysis (Get started with computer vision in Azure) (Microsoft Learn)

  3. Question 3

    A finance team wants to paste screenshots of sales charts into a chat app and ask questions such as "Which region grew fastest?". Which kind of model should the app use?

    • A An image generation model such as GPT-Image-1
    • B A vision-enabled GPT model
    • C An embedding model
    • D A text-to-video model such as Sora
    Show answer and explanation

    Answer: B

    Vision-enabled GPT models accept images in prompts and can answer questions about them, including extracting meaning from charts and screenshots.

    • A: Image generation models create new images. They don't answer questions about the images you give them.
    • C: Embedding models turn content into vectors for search. They don't answer questions.
    • D: Sora generates video from text. It doesn't analyze charts.

    Source: Multimodal models for image analysis (Get started with computer vision in Azure) (Microsoft Learn)

  4. Question 4

    Before writing any code, you want to see how a deployed model interprets some sample product photos by uploading them and trying different prompts. Where should you do this?

    • A In the model playground in the Foundry portal, using a vision-enabled model
    • B In the model catalog, on the model's benchmark page
    • C In a Global Batch deployment
    • D In Azure Content Understanding, with the prebuilt-imageSearch analyzer
    Show answer and explanation

    Answer: A

    In the Foundry portal, the model playground lets you chat with a deployed model. With a vision-enabled model you can upload images and test prompts interactively.

    • B: Benchmarks compare models on standard tests. You can't upload your own photos there.
    • C: Batch deployments process large jobs over hours, not interactive tests.
    • D: Content Understanding runs its own analyzers. It wouldn't show how your deployed model responds to your prompts.

    Source: Multimodal models for image analysis (Get started with computer vision in Azure) (Microsoft Learn)

  5. Question 5

    A marketing team wants to create new product images from text descriptions in Microsoft Foundry. Which kind of model should it deploy?

    • A Sora
    • B A vision-enabled GPT chat model
    • C A GPT-Image series model
    • D text-embedding-3-small
    Show answer and explanation

    Answer: C

    GPT-Image series models generate (and edit) images from text descriptions. They replace the older DALL-E models, which have been retired.

    • A: Sora generates video, not still images.
    • B: Vision-enabled chat models understand images you send them, but they answer with text.
    • D: Embedding models turn text into vectors. They don't create images.

    Source: How to use image generation models (Microsoft Foundry) (Microsoft Learn)

  6. Question 6

    A team needs an 8-second product video created from a text prompt, using a model provided directly through Microsoft Foundry. Which model should they use?

    • A Sora 2
    • B GPT-Image-1
    • C A vision-enabled GPT model
    • D Content Understanding's prebuilt-videoSearch analyzer
    Show answer and explanation

    Answer: A

    Sora models are currently the only native video generation models provided directly through Foundry. Sora 2 also adds audio generation, improved realism and remixing.

    • B: GPT-Image-1 creates still images, not video.
    • C: Vision-enabled models can understand images, but they don't generate video.
    • D: prebuilt-videoSearch analyzes existing videos. It doesn't create new ones.

    Source: Video generation models (Get started with computer vision in Azure) (Microsoft Learn)

  7. Question 7

    Your app generates short videos with a Sora model in Microsoft Foundry. For each statement, choose Yes if it is true or No if it is false.

    • Video generation usually runs as an asynchronous job: you create it, check its status, then download the video.
    • A video is usually returned within a second of the request.
    • Sora models include responsible AI restrictions, such as limits on generating real people.
    Show answer and explanation
    Video generation usually runs as an asynchronous job: you create it, check its status, then download the video.Yes
    A video is usually returned within a second of the request.No
    Sora models include responsible AI restrictions, such as limits on generating real people.Yes

    Video generation is resource-intensive, so it runs as an asynchronous job that often takes 1–5 minutes: create the job, poll its status, then download the result. Sora models also include responsible AI restrictions on things like real people and copyrighted characters.

    Source: Video generation models (Get started with computer vision in Azure) (Microsoft Learn)

  8. Question 8

    A blog platform wants to generate thousands of simple thumbnail images, and cost matters more than top quality. Which model is the best fit?

    • A GPT-Image-2.5-Sunburst
    • B GPT-Image-1-Mini
    • C Sora 2
    • D A vision-enabled GPT model
    Show answer and explanation

    Answer: B

    GPT-Image-1-Mini is a lighter-weight, more cost-efficient version of GPT-Image-1. Microsoft describes it as best for fast prototyping, bulk generation and cost-sensitive use cases.

    • A: GPT-Image-2.5-Sunburst is the most capable model, best when editing precision matters. That isn't the priority here.
    • C: Sora 2 generates video, which is far more expensive and not what's needed.
    • D: A vision-enabled chat model understands images but doesn't generate them.

    Source: How to use image generation models (Microsoft Foundry) (Microsoft Learn)

  9. Question 9

    Your Python app uses the Responses API in Microsoft Foundry, and you want the response to include an image generated by your GPT-Image deployment. What must the request include?

    • A An input_image content part
    • B A tools list containing the image_generation tool
    • C A higher max_output_tokens value
    • D An audio modality setting
    Show answer and explanation

    Answer: B

    In the Responses API, you ask for an image by including the image_generation tool. The response then contains an image generation result with the image as base64 data, which the app decodes and saves. (In Foundry, the request also names your image deployment, for example in the x-ms-oai-image-generation-deployment header.)

    • A: input_image sends an image to the model as input. It's used when the model should look at a picture, not create one.
    • C: More output tokens allow longer text. They don't switch on image generation.
    • D: Audio settings are for spoken input and output.

    Source: Image generation models (Get started with computer vision in Azure) (Microsoft Learn)

  10. Question 10

    A travel company has a beach photo it likes, but wants a GPT-Image model to replace only the cloudy sky with a sunset and leave everything else untouched. What should it use?

    • A The image generation API with a more detailed prompt describing the whole scene
    • B A vision-enabled GPT model, sent the photo and the instruction
    • C Sora 2 remix
    • D The image edit API with a mask that marks the sky as the area to change
    Show answer and explanation

    Answer: D

    The image edit API changes an input image based on a prompt. A mask, a PNG the same size as the photo, marks with transparent pixels exactly which part to edit, so the rest of the photo stays as it is.

    • A: The generation API creates a brand-new image, so the beach and everything else would change too.
    • B: A vision-enabled GPT model can describe the photo, but it doesn't return an edited image.
    • C: Sora 2 remix edits videos, not still photos.

    Source: How to use image generation models from OpenAI (Microsoft Foundry) (Microsoft Learn)

  11. Question 11

    Your app calls the image generation API for a GPT-Image-1 series deployment in Microsoft Foundry. For each statement, choose Yes if it is true or No if it is false.

    • The API returns the images as base64-encoded data rather than as URLs.
    • If the prompt is flagged as harmful, the API still creates an image but blurs it.
    • One call can return more than one image.
    Show answer and explanation
    The API returns the images as base64-encoded data rather than as URLs.Yes
    If the prompt is flagged as harmful, the API still creates an image but blurs it.No
    One call can return more than one image.Yes

    GPT-Image models return base64-encoded image data, which your app decodes and saves. You can ask for between 1 and 10 images in one call. Prompts go through a content filter, and a flagged prompt returns an error instead of an image.

    Source: How to use image generation models from OpenAI (Microsoft Foundry) (Microsoft Learn)

  12. Question 12

    In Microsoft Foundry, you deploy the gpt-4.1-mini model with the deployment name "photo-helper". Your Python app sends an image and a question with client.responses.create(). What value should you pass as the model argument?

    • A photo-helper
    • B gpt-4.1-mini
    • C The Foundry project's name
    • D The resource endpoint URL
    Show answer and explanation

    Answer: A

    When you call a model you deployed in Foundry, the model argument takes your deployment name. The deployment is what your app calls, and it can have a different name from the underlying model.

    • B: gpt-4.1-mini is the model's name in the catalog. Requests go to your deployment, which you named photo-helper.
    • C: The project name identifies your workspace, not the deployment that should answer the request.
    • D: The endpoint URL is used to create the client (base_url), not as the model argument.

    Source: Multimodal models for image analysis (Get started with computer vision in Azure) (Microsoft Learn)

Want more practice on Vision and image generation?

Practice the full question bank for this area, in random order, with your progress saved in your browser.

Practice Vision and image generation