You've made AI write and AI draw. Today it does something wilder: it looks at a picture and writes a sentence about it — out loud, in words. Draw literally anything, hit describe, and a real AI on your own laptop will tell you what it sees.
Scribble a cat, a house, a spaceship — whatever. Then press 🔮 Describe it. A real image-captioning model (running 100% on your device) will look at the pixels and generate a sentence. No lookup, no cheating — it reads the picture and invents words to match.
The photo buttons need internet the first time. The drawn scenes and smiley work totally offline. 🌐
Make a random rainbow mess — pure abstract scribble that means nothing. Then describe it. The model will never say "uh, I don't know." It always produces a confident sentence, even when there's nothing real there.
The AI captions a picture. Your job: is the caption a good match or wrong? If it's wrong, pick the caption that actually fits. Judge 3 pictures correctly to earn the badge. This is a real job — humans grade AI captions to make them better!
Captioning isn't a toy — it quietly runs inside apps you use. Same core move every time: a machine looks, then writes.
Blind users browse with a screen reader. Captioning writes the "alt text" so they can hear what every photo shows. Genuinely life-changing tech.
Search your camera roll for "dog on a beach" with no tags. The photos got captioned, so now words can find them.
Sites auto-describe billions of uploads to flag ones that break the rules — far faster than any human team could.
Everything above already used a real model — this is the same one, pointed at real internet photographs so you can see how it does on the hard stuff. It's ViT-GPT2, running 100% on your machine. Load it once, then describe any photo.
First load pulls the model from Hugging Face, then it's cached for offline use. On a 4GB laptop give it a minute. ☕
You ran a real multimodal AI: a vision encoder read your pixels, a text decoder generated a sentence. You also caught it hallucinating — sounding sure about nonsense — the one thing to always watch for. That's exactly how alt-text, visual search, and every VLM work.