How to Use an AI Image Describer: Real Outputs From All Seven Modes

A batch of product photos goes live tomorrow, and every one still needs alt text, the short description a screen reader reads out loud. A few photos in, the descriptions start sounding the same.

AI can write that first draft for you in seconds. What it can't do is decide which kind of description the job needs, or which lines to double-check before they go live. So we ran all seven modes of the Vormly AI image describer on real images and kept every answer, mistakes included.

TL;DR

Writing alt text or a caption? Start with Describe Briefly. Need every detail written down? Use Describe in Detail. Need the words printed in an image? Use Extract Text. Need one fact? Use Ask a Question. Whichever you pick, read the answer once before you publish. The describer is good at saying what's in a picture, and less reliable when it guesses, when it counts lots of similar things, or when it lists one object as two.

What an image describer does for you

Picture a folder of images that all need words: a caption for each post, a description for each product, the text from a pile of scanned receipts. Writing it all by hand is slow, and it gets sloppy by the end.

This kind of tool does that writing for you. You upload a picture, an AI that can read images looks at it, and it answers in plain text. Depending on what you ask for, you get a short description, a long one, a list of what's in the picture, the words printed in it, or an answer to a question.

People use it for jobs like these:

• Alt text and captions. People who use screen readers hear a short description instead of seeing the image. The web's accessibility rules, in WCAG Success Criterion 1.1.1 (opens in a new tab), ask for one on every image that matters to the page.
• Keeping big image collections searchable. A product library or photo archive is much easier to search when every picture has a description.
• Copying text out of images. Signs, screenshots, posters and scans turn into text you can paste, instead of typing it all again.
• Quick questions. Ask something specific, like what a label says, and get a straight answer.

How we tested

This is the describer: upload an image on the left, pick one of the seven modes and the language you want, and the answer appears on the right. It takes PNG, JPG and WebP files up to 10 MB, it's free, and you don't need an account.

For each mode, we picked an image from the Vormly gallery that suits that mode's job and ran it on September 15, 2026. Before the AI looks at an image, it's resized so the longest side is 1,024 pixels, exactly as it would be for you. The AI words things a little differently each time, so treat the mistakes below as the kinds of things to watch for, not an exact list.

The model that reads the image

Behind every mode is Google's Gemini 2.5 Flash-Lite (opens in a new tab), an AI model that can take in text, images, video, audio and PDFs and answer in text. Google built it to be fast and cheap for simple jobs done at large scale, like sorting things into groups or pulling out basic information. We use the stable version released on July 22, 2025, and Google's Gemini deprecation schedule (opens in a new tab) doesn't list a date for switching it off.

That explains two things you'll see below. The seven modes aren't seven different AIs: each one sends your image to the same model with different instructions. And the model is at its best with plain facts, like which objects are there, what text is written, or how many of something there are when there are only a few. It's weaker when it starts guessing, like the time of day in a photo or which art movement a painting leans toward.

Which mode to use for which job

Describe in Detail: thorough, but check its guesses

Say you're building a catalogue or an archive and need a proper written record of each image, not just a caption. That's what this mode is for.

It noticed nearly everything that's really there: the open book, the stack of old books with a key beside it, the box and the inkwell with a quill, the window, the dust floating in the light, and a low camera angle, as if you were sitting at the desk.

The mistakes are in the sentences where it guesses, and they sound just as sure as the rest:

• It can't decide about the book. It says the open book suggests it "is currently being read or has just been closed." A book lying open hasn't just been closed.
• It guesses the time. It calls the scene "likely late morning or early afternoon," judging from the light. Nothing in the picture tells you the time.

Describe Briefly: the starting point for alt text

This is the mode for that batch of product photos, or any image that needs a caption or alt text.

It's capped at 40 words, and here it used 25: what the photo shows, the one detail you'd notice first (the green eyes), where the cat is, and what's behind it. Every word of it is accurate.

Treat it as a draft anyway. If your page is about the cat, you can drop the sofa and the plant. And if the photo already has a caption that says the same thing, the alt text can be left empty. The alt text section near the end explains when.

Describe the Person: what someone looks like, never who they are

This one helps when portraits or team photos need descriptions, for a website or a photo library.

It stuck to what you can see: hair, glasses, expression, the tilt of the head, clothes and how the photo is framed, and it got all of it right. It's built never to say who a person is, even if they look like someone famous. If there's no one in the picture, it says so and describes what is there.

The one thing to handle carefully is the age: "likely in her late teens or early twenties." That's a guess, and not something to publish about a real person.

Recognize Objects: a tidy list, with one item counted twice

Say you sell clothes online and want a list of everything in each product shot. This mode gives you that list.

The list is neat, grouped the way a shop would group it, and easy to paste into a spreadsheet. It even spotted the small black label inside the jacket collar.

It also made one mistake. It lists a "Green scrunchie," but that's the bag's ruffled handle, so the list has one more item than the photo does. This mode is good at spotting what's there, and less reliable at telling where one object ends and the next begins.

If you're describing clothes in particular, the outfit describer goes further: each garment, its fabric, and the style they add up to.

Analyze Art Style: good at naming a look, not at how it was made

Found an image you love but can't put a name to its style? This mode helps, whether you're briefing a designer or searching for more images like it.

The useful part is the vocabulary. It recognized Cubism right away, mentioned Pablo Picasso and Georges Braque, and explained how the picture is broken up into flat, angular shapes. That's the language you need to describe a look to someone else.

Two parts are less reliable. It says the picture is an oil painting, "evident from the visible texture," but it's an AI image made to look painted. And it links the style to Orphism because of its "vibrant, faceted planes and a focus on color," then a paragraph later describes the palette as "earthy and subdued." Take anything it says about how a picture was made as a guess, and watch for answers that contradict themselves.

Extract Text: the words come out right, the notes about where don't

When you need the words from a sign, a screenshot or a poster, retyping them is slow and easy to get wrong. This mode copies them out for you.

Every word came out exactly right: "A-83" all three times it appears, and コンピューター kept in Japanese, not translated. The only mistake is the last line, which lists a red rectangle that has no text in it.

Changing the output language works a little differently here. When we picked Spanish, the notes about where each piece of text sits came back in Spanish, but the Japanese stayed Japanese. That's what you want when you're copying text rather than translating it:

This run also shows the mode's weak spot. It listed the label on the computer's base twice, and said the monitor's label is at the bottom left, which it isn't. The words are dependable. The notes about where they sit are not.

Two more things to know. If part of the text can't be read, it's marked [illegible], and an image with no text comes back as "No text detected in this image." And because images are shrunk to 1,024 pixels before the AI reads them, very small text can be missed. When we tested a website screenshot earlier, titles inside tiny thumbnails were skipped until we cropped the image down to that area.

Ask a Question: one fact at a time

Sometimes you don't need a description at all, just one answer: what a label says, or whether something is in the picture.

We asked two questions. The image could answer one of them, but not the other.

Correct: three whole lemons, plus one cut in half, which the question said to leave out.

Also correct. Nothing in the picture shows where the lemons came from, so it says it can't tell and describes what it can see instead of making something up.

Be careful with counting, though. Three big lemons are easy. In an earlier test, on an image full of overlapping yellow circles, the same kind of question got the answer 10 when only two circles were fully visible.

Alt text: when you don't need the describer at all

Back to that batch of product photos. Before you create alt text for every image, it helps to know that some images shouldn't get a description at all.

Alt text comes from an accessibility rule, WCAG Success Criterion 1.1.1 (opens in a new tab): every image that matters needs a text version that does the same job as the image. What matters is the job, not the picture. The W3C's Images Tutorial (opens in a new tab) and its alt text decision tree (opens in a new tab) start from what an image is for, not what's in it. That tells you which mode to use, and sometimes that you don't need one:

If you wanted a prompt, not a description

A description is written for people to read. If what you actually want is text you can give an AI image generator to recreate or remix a picture, that's a different job, and there's a different tool for it.

Conclusion

Back to that batch of product photos due tomorrow. You don't have to write every description from scratch. Let the describer write the first draft, pick the mode that fits the job, and save your time for a quick read of each answer: look for guesses, items counted twice, and notes about where things sit.

Try the AI image describer on your own images. It's free, and you don't need an account.