Microsoft AI-103: Implement computer vision solutions Practice Questions

Prompt-driven, mask-based editing constrains the change to the selected shirt region, preserving pose, face, and background exactly as required. Option A risks regenerating the entire image, which would not guarantee the pose and background stay unchanged. Option B is the wrong modality — this is a still-image editing task, not video generation. Option D only locates the shirt but performs no edit, so the requirement is not met.

Inpainting works on an existing image regardless of its origin — a previously generated image can serve as the base, with a mask isolating just the region needing refinement (like a distorted hand), leaving the rest untouched. Option B is a false restriction; inpainting is not limited to photographs. Option C is a valid but wasteful alternative that this workflow specifically avoids, and does not guarantee the rest of the image stays the same. Option D introduces an unnecessary and incorrect video conversion step.

Visual question-answering grounded in the image directly supports answering specific, open-ended user questions about visual evidence in that image, such as identifying a broken part. Option A produces a generic overview caption and is not designed to answer targeted questions. Option C is a generative capability aimed at creating new imagery, not answering a diagnostic question about the existing photo. Option D targets accessibility description text, not free-form Q&A grounded in the pixels.

Mask-based inpainting lets you constrain edits to a selected region (the background) while leaving unmasked pixels — the product — untouched, which is exactly the requirement. Option A (full regeneration) cannot guarantee the product remains pixel-identical since the whole image is resynthesized from the prompt. Option C is the wrong modality entirely; video generation does not solve a still-image editing task. Option B abandons the generative service and produces only a text description, not an edited image.

Visual policy rules that specifically check for watermarks, prohibited symbols, and brand usage are the correct control — this is a distinct governance layer from general harm-category safety filtering, since the images in question are not unsafe, just non-compliant with brand policy. Option A only targets harm categories like violence or adult content and would not catch a competitor logo. Option C is unrelated to image content. Option D provides a description but enforces nothing.