Captioning

Captioning is the AI generation of a natural-language text description of visual content — a still image or a video — used most commonly for accessibility (screen-reader alt text), content indexing/search, and automated metadata generation. Note this is distinct from subtitle captioning (timed on-screen text of spoken dialogue, which overlaps with transcription); "image captioning" in the AI/ML sense refers to a vision model describing what's depicted in an image ("a golden retriever catching a frisbee in a park"). Modern captioning is handled by vision-language models — multimodal LLMs (GPT-4o, Claude, Gemini) or dedicated models (BLIP-2, LLaVA) trained on large paired image-text datasets — that jointly encode the image and generate a fluent description, and can be prompted for different levels of detail or specific focuses ("describe this for a blind user, focusing on function over aesthetics" vs. "write an SEO-optimized alt attribute in under 125 characters"), giving builders a single flexible API rather than needing separate specialized models for each captioning use case. Why it matters for SaaS builders: automated captioning is required at scale for any platform with significant user-generated visual content — e-commerce (auto-generating alt text for thousands of product photos, both for accessibility compliance like WCAG and for image-search SEO), stock-photo/DAM platforms (auto-tagging and describing assets for searchability), and social/content platforms serving visually impaired users. It's a low-cost, high-compliance-value feature since manually writing alt text for a large media library is tedious and frequently skipped. A concrete worked example — an e-commerce platform auto-generating alt text at scale: (1) on product-image upload, a background job sends the image to a vision-language model with the prompt "Write a concise, descriptive alt-text attribute for this product photo in under 125 characters, suitable for accessibility and SEO. Do not start with 'image of.' Mention material, color, and key visual details."; (2) the model returns "Matte black ceramic pour-over coffee dripper with wooden collar, on white background"; (3) the alt text is saved to the product's `image_alt` field and rendered in the `<img alt="...">` tag, immediately improving both screen-reader accessibility for visually impaired shoppers and eligibility for Google Image Search traffic; (4) a nightly batch job backfills alt text for the previously empty legacy product catalog, processing a fixed batch size per run to control API cost, and flags any image where the model's confidence is low (e.g., an ambiguous lifestyle photo rather than a clean product shot) for a human to write manually instead.