Implement Computer Vision Solutions · Topic 1 of 2
Image and Video Generation
AI-103 media generation: images from text and reference media, inpainting and mask edits, Sora 2 video jobs, parameters such as seconds and size, remixing.
- 4 min
- Image generation: text to image, reference images, mask-based inpainting, prompt-driven edits.
- Video generation is an asynchronous job: submit, poll, download.
- Sora 2: seconds 4, 8 or 12; size 1280x720 or 720x1280; input_reference; remix_video_id.
What it is
Foundry offers image generation models that create pictures from text prompts and reference images, and video generation models such as Sora 2 that create short clips from text and reference media. AI-103 tests how to call them, how to edit their output, and which controls shape the result.
How it works
Image generation
| Task | How it works |
|---|---|
| Text to image | Send a prompt describing the subject, style and composition |
| Image from reference media | Supply one or more input images to guide style or content |
| Inpainting (mask-based edit) | Supply the image and a mask, a PNG the same size as the image; the fully transparent area of the mask marks where the model may change the image |
| Prompt-driven modification | Describe the change to make to an existing image |
Typical controls include the output size, quality level, the number of images and the output format. Prompts work best when they are specific about subject, setting, style, lighting and what to avoid.
Video generation with Sora 2
Video generation is an asynchronous job: you submit a request, poll the job status, then download the result.
| Property | Meaning |
|---|---|
| prompt | Text describing the clip |
| seconds | Clip length; Sora 2 accepts 4, 8 or 12 |
| size | Resolution as a single value: 1280x720 (landscape) or 720x1280 (portrait, the default) |
| input_reference | A single reference image used as the anchor for the first frame; it must match the requested size exactly |
| remix_video_id | Starts from an existing generated video to edit it with a new prompt |
The earlier Sora preview used different parameter names (such as width, height, n_seconds and n_variants). Questions on AI-103 follow the current API.
A worked example: a studio wants an 8-second landscape clip that starts from an approved product shot. The request sets the prompt, seconds to 8, size to 1280x720 and input_reference to the product image at exactly 1280x720, then polls the job until it completes.
Editing generated video
- Remix an existing video by referencing it and describing the change, for example a different time of day.
- Limit each remix to one clearly described change; narrow edits keep more of the original video.
- To compare alternative takes, submit separate generation jobs and choose the best one.
Choosing generation and editing controls
| Requirement | Control |
|---|---|
| Keep a brand character consistent | Supply reference images |
| Change only part of an image | Inpainting with a mask |
| Portrait clip for mobile | size 720x1280 |
| Several image options to choose from | Number of images (n) |
| Change an existing clip | Remix from its video ID |
In a mask edit, the masked (transparent) region is what changes. Everything outside it is kept.
How the exam asks about this
- "Replace only the sky in a product photo." Inpainting with a mask over the sky.
- "A clip must start from an approved product image." Set input_reference to that image, at the same size as the video.
- "Change the lighting of a clip already generated." Remix it using its video ID.
What to remember for the exam
- Image generation: text to image, reference images, mask-based inpainting, prompt-driven edits.
- Video generation is an asynchronous job: submit, poll, download.
- Sora 2: seconds 4, 8 or 12; size 1280x720 or 720x1280; input_reference; remix_video_id.
- Old preview names (width, height, n_seconds, n_variants) are not the current API.
- Choose controls that match the requirement: masks for partial edits, references for consistency.
Video generation does not return the clip in the first response. The app must poll the job and fetch the output when it is ready.
Common mistakes
Using separate width and height for Sora 2. Use a single size value.
Asking for a 10-second Sora 2 clip. Allowed lengths are 4, 8 or 12 seconds.
Regenerating a whole image to change one region. Use a mask edit.
Check yourself
Which Sora 2 property anchors the first frame to an image you supply?
input_reference, an image that must match the requested size.
What does the mask do in an image edit?
Its transparent area marks the region the model may change; the rest is kept.
How do you change an existing generated video?
Remix it by referencing its video ID with a new prompt.
Topic 1 of 2 in Implement Computer Vision Solutions
Test yourself on Implement Computer Vision Solutions
The full Microsoft AI-103 question bank has 648 questions, with timed practice exams and an explanation for every answer.
This study guide is independent and is not affiliated with, endorsed by, or sponsored by Microsoft. Vendor names, certification names and exam codes are trademarks of their respective owners and are used for identification only.