Image-to-Text Generation (Captioning)

Produces prose describing what an image contains. The dominant production use is accessibility — alt text at scale — with search indexing and content moderation close behind.

Large image libraries needing accessible descriptions or searchable text, where writing them by hand is not viable.

Where the image content is consequential and specific. A caption is a summary and will omit whatever the model judged unimportant, which may be the thing that mattered.

Captions entering a public product need bias and safety filtering, and where a caption becomes a record it needs the same provenance marking as any other generated content.

Fluent vagueness. The caption reads well, describes the scene broadly, and misses the specific detail that made the image worth captioning.

A retailer generating alt text for several hundred thousand product images so the catalogue is navigable by screen reader.