What this is. This is a demo, not a finished product. Argira is an internal prototype exploring whether it's possible to generate alternative text for images automatically, without relying on an AI model trained to describe images.
How it works. When analyzing an image, the tool draws it onto a hidden canvas and measures statistical properties of its pixels — color variance, edge density, hue and saturation entropy. Using those metrics, a heuristic (a set of weighted rules, not a trained model) estimates whether the image is probably a photograph, a painting, an illustration, or an AI-generated image. Based on that type, an alternative-text sentence is built from a simple template. Unlike most tools of this kind, Argira exposes the metrics and the confidence behind the result instead of delivering a final sentence without showing the uncertainty underneath.
What this demo does not do. This particular demo does not perform semantic recognition via a vision model: it cannot identify objects, people, or scenes — it cannot say "a dog running on the beach," only "this looks like a photograph." The broader ARGIRA project includes other tools capable of working with an image's actual content (available at argira.eus), but this particular demo uses only the statistical heuristic and templates described above.
Why some images can't be analyzed. Due to a browser security restriction, pixel content can only be read from images whose origin server explicitly allows it, or from files uploaded locally. This is not a flaw in the tool, but a restriction of the browser itself.
How this compares to other tools. Commercial tools already exist (AltText.ai, Alt Magic, alt-text generators built into some CMSs) that use trained vision models to describe an image's actual content — "a blue wool sweater," "a golden retriever playing on the grass." This demo doesn't compete with that or try to replace it: it explores a different question — whether, without a vision model, it's at least possible to classify the type of image while being transparent about the uncertainty of that estimate.
A heuristic based on pixel statistics will never recognize objects or scenes, but it can complement a vision model by contributing context about its own certainty — for example, if it classifies an image as "possibly synthetic" or reports low confidence, that information can accompany the alternative text generated by another system. This matters for accessibility: screen-reader users often have no other way to judge the reliability of an automatically generated description.