Captions matched to your photo

It matches your photo to a topic. It does not describe the photo. An image model called CLIP scores how well a picture and a short line of text fit together, and the tool uses that to compare your photo with a one-line description of every topic we have captions for. It has no idea whose birthday it is or which beach you’re on.

• It can’t read text in the photo, such as a poster, a menu or a screenshot.
• It can’t tell who is in the picture or what the occasion is, unless the photo shows it plainly, like a cake with candles.
• It only knows the topics we have captions for. If yours isn’t one of them, search for the nearest.
• The captions aren’t written for your photo. They are lines from our caption pages, so everyone who picks the same topic sees the same ones.

We ran the tool on a handful of photos in October 2026. This is what it switched on. WebGPU and the slower WebAssembly fallback gave the same result.

The misses are as useful as the hits. The serum photo has a phone in it, and Mirror selfie came first, when the topic you’d want is Product launch, one tap away. In our tests, photos with a person in them often brought up Outfit of the day, Mirror selfie or Selfie. Treat the topics as a first guess you can change.

Photos are often two things at once: a dog on a beach, friends at a birthday dinner. Switch on two or three topics and the captions take turns between them, best match first, so the list isn’t all from one page. “Kind of caption” narrows it to one situation, such as “Dogs at the beach” in the Beach topic.

Nineteen of the 153 topics are about a mood, a style of post or a relationship rather than something you can see: Funny, Savage, Cute, Girlfriend, Photo dump and so on. The model leaves those out of the matching. Open “Switch or add a topic”, choose Moods and styles, and add one to tilt the list towards that tone.

1. Pick the caption closest to what you want to say, even if it isn’t perfect.
2. Change one detail so it is true of your post: the name, the place, the number.
3. Read it aloud once. If it sounds like something you’d say, post it.

Most of the captions are one or two short sentences, so there is room to add a place, a name or a second line. Meta’s reference for publishing Instagram posts through its API lists a maximum of 2,200 characters for a caption (we read it in October 2026). Our caption length guide covers what Meta does and doesn’t document.

Your photo is read in your browser and never uploaded. The first photo needs two downloads: the CLIP image model, about 64 MB from Hugging Face, and about 4 MB of code to run it, from jsDelivr. Your browser keeps them, so the next photo starts at once. The text half of CLIP was run ahead of time and its results ship with this page, so there is no second model to download.

After the downloads, a photo took about a second on a laptop, with WebGPU or without. Phones and older laptops will be slower. Picking a topic by words or by browsing needs no model at all.