WritingmateWritingmate

Image Caption Generator: How to Write Captions and Alt Text From a Photo (Free Tool, Prompt Template, and a 3-Minute Accuracy Check)

Most image caption generator pages say upload a photo and click. This guide sorts caption jobs by type, gives a prompt template that pins platform, tone and what is visible, and ends with a 3-minute check for invented details.

Try the free image caption generator
200+ models
One subscription
No API keys
Cancel anytime
A photo of a coffee shop table next to a generated caption and an alt text field with a checklist beside it
Artem Vysotsky

Author, Co-Founder & CEO

Artem Vysotsky

Sergey Vysotsky

Reviewer, Co-Founder & CMO

Sergey Vysotsky

9 min read
Updated: 11/22/2026

Most image caption generator pages tell you the same thing: upload a photo, click a button, copy the result. Then you post it and realize the AI called your friend's dog a "golden retriever puppy at sunset" when it was a ten-year-old lab at 2 p.m. in a parking lot.

My name is Artem, and I run the Writingmate blog. I've spent a lot of hours feeding photos to vision models and checking what comes back, and the pattern is consistent: the captions read nicely, but the details are where they go wrong. This guide shows you how to sort your caption job, use a prompt that pins down what the model is allowed to say, and run a 3-minute check before anything gets posted.

You can follow along with the free Writingmate image caption generator, or paste the same prompt into any chat that accepts images.

How I compared caption jobs

I don't think "image caption" is one task. It's at least four, and they fail in different ways. Here's what I checked for each one, using a handful of my own photos (a café table, a street scene with a shop sign, a product shot on a white background, and a group photo):

  • Did it describe what's visible? Colors, objects, setting, number of people.
  • Did it invent anything? Names, locations, brands, text on signs, time of day.
  • Does it fit the platform? Length, tone, hashtags or no hashtags.
  • Would I post it with one edit or fewer?

That last one is the real test. A caption that needs a full rewrite isn't saving you time. A note on method: these are my own observations from a small set of photos, not a benchmark. Treat them as a starting point and run your own images through the same steps.

Sort your caption job first

The same photo needs very different text depending on where it's going. Here's the split I use:

Job

What it needs

Typical length

Main risk

Instagram caption

A hook, a personal line, optional hashtags

1-3 short lines before the "more" cut

Generic filler, invented location

TikTok caption

Punchy, conversational, a reason to watch

One line plus a few tags

Claims the visual doesn't support

Alt text

Plain description of what's in the image and why it matters

Roughly one or two sentences

Guessing at identity, emotion or text

Product photo

Visible attributes: color, material, angle

One sentence for alt, a few for listing copy

Specs and materials that aren't visible

Notice that alt text and social captions are almost opposites. A social caption sells a feeling. Alt text describes the image for someone who can't see it. If you ask one prompt to do both, you usually get a mushy result that fails at both.

The prompt template that pins down what's visible

Here's the thing: vision models are happy to fill gaps. If you give them a photo and say "write an Instagram caption," they'll add a story. The fix is to tell them exactly what they can use. This is the template I paste in:

Look at the attached photo. Write [number] caption options for [platform: Instagram / TikTok / alt text / product listing].
Tone: [casual / funny / professional / calm].
Length: [e.g. under 150 characters].
Only describe what is clearly visible. Do not name people, places, brands, or events unless I give them to you below. If there is text in the image, quote it exactly or say it's unreadable.
Facts I'm supplying: [location, names, occasion, or "none"].
After the captions, list anything you were unsure about.

Three lines in there do most of the work:

  • "Only describe what is clearly visible." This cuts down on the invented backstory.
  • "Facts I'm supplying." You give it the real location and names, so it doesn't guess.
  • "List anything you were unsure about." This gives you a shortlist of things to check, which makes the final review much faster.
Screenshot of the Writingmate image caption generator with a photo uploaded and caption options generated

In the free tool, you upload the image and get captions to start from. If you want tighter control, open the same photo in Writingmate chat and paste the template above. There you can also switch between vision-capable models and compare how each one handles the same photo. You can see what's available on the models page.

Instagram and TikTok: how to get a caption that sounds like you

For social posts, the biggest problem isn't accuracy, it's sameness. Ask any model for an Instagram caption and you'll get "Coffee and good vibes ☕✨". Here's how I get away from that:

  1. Give it a voice sample. Paste two of your old captions and say "match this tone."
  2. Ask for different angles. One funny, one sincere, one that's just a question.
  3. Cap the hashtags. Say "no more than 3 hashtags" or "none." Otherwise you get a wall of them.
  4. Cut the first line yourself. Models love an opening like "Just another day in paradise." Delete it and start with the second sentence.

For TikTok, add one more constraint: "the caption should make someone want to watch the video, not summarize it." Short, a little curious, and no claims about things that aren't on screen. If your clip shows a cooking step, say so. Don't let the model add "the secret ingredient" unless there is one.

I also found it useful to generate five options and mix lines from two of them. It takes an extra minute and the result sounds much less like a machine wrote it.

Alt text: different rules, different prompt

Alt text is where mistakes matter most, because a screen reader user takes the description at face value. The American Foundation for the Blind makes this point directly in its piece on alt text in the age of AI: a person may trust an AI-generated description, act on it, and only later find out it was misleading or wrong.

Accessibility researcher Adrian Roselli put the concern bluntly on X:

"My attempt to use AI to demonstrate the problems with relying on AI." — @aardrian on X

That's why I use a separate alt text prompt:

Write alt text for this image. One or two sentences, under 125 characters if possible. Describe the subject and setting plainly. Don't start with "image of" or "photo of." Don't guess at names, emotions or intent. If the image contains text that matters, include it exactly. If the image is purely decorative, reply "decorative."

A few rules of thumb I follow:

  • Describe purpose, not every pixel. A photo in a blog post about hiking boots needs "Brown leather hiking boots on a rocky trail," not a list of every rock.
  • Include text that's in the image. A shop sign or a chart label can be the whole point.
  • Skip the guesses about feelings. "Looks happy" is a guess. "Smiling" is visible.
  • Let the model say "decorative." Not every image needs a description, and a page full of filler is worse for screen reader users than silence.

Plenty of university and accessibility groups say the same thing in different words: use AI as a drafting tool, then have a person review it. James Madison University's accessibility team frames AI-generated alt text exactly that way.

Product photos: describe what you can see, not what you know

This is where I've seen the most confident nonsense. You upload a photo of a water bottle and the model writes "double-wall vacuum insulated stainless steel, keeps drinks cold for 24 hours." Nothing in the image says that. The model is pattern-matching to similar listings.

My fix is simple. Put the real specs in the prompt under "Facts I'm supplying," and tell the model that everything else must come from the image alone. If you're writing full listings rather than captions, I covered that workflow in the product description guide.

Side-by-side of a product photo with an AI caption highlighting an invented spec in red and verified details in green

The 3-minute accuracy check

Before you post or publish, run this. It really does take about three minutes once you've done it a couple of times.

Minute

What to check

What you're looking for

1

People

Any names, ages, relationships or job titles the model added. Remove any you didn't supply.

2

Places and time

City names, landmarks, "at sunset," "in summer." Check against the photo and your own memory.

3

Text and numbers

Zoom in on every sign, label or screen. Compare the quoted text to the real thing, letter by letter.

Then one last pass: read the "things I was unsure about" list from the prompt. Every item on it is a place the model told you it was guessing. Fix or delete those lines.

Here's what typically trips up the check:

  • Misread text. Signs in a different alphabet or at an angle get "translated" into something plausible but wrong.
  • Wrong counts. "Three friends" when there are four, especially if someone's half out of frame.
  • Invented brands. A generic sneaker becomes a specific brand.
  • Emotion guesses. "Excited," "nostalgic," "peaceful." You can keep these in a social caption if they're true, but never in alt text.

A Ministry of Testing forum thread about using ChatGPT for alt text echoes this. The poster found it worked well overall but suspected some output was guesswork, like a misread expression and mixed-up speakers in a meme. You can read it in the original discussion.

Where the free tool fits, and when to use chat instead

I use the free generator when I want a quick start on a single photo: upload, get options, edit. It's fast and there's no setup. I switch to chat when I need one of these:

  • A specific voice (pasting my past captions)
  • Several photos with consistent wording, such as a product series
  • A second opinion from a different vision model on the same photo

If you end up doing this daily, check the pricing page to see which plan covers the usage you need.

So here's the bottom line

A caption generator is a good first draft and a bad final answer. Sort the job first, because Instagram, TikTok, alt text and product shots all want different things. Use a prompt that limits the model to what's visible and gives it your real facts. Then spend three minutes checking people, places and text. That's the difference between a caption that sounds good and one that's true.

Try it on one of your own photos and see what the "unsure" list tells you. You'll probably be surprised how much the model flagged on its own.

See you in the next one!

Artem

Frequently Asked Questions

Artem Vysotsky

Written by

Artem Vysotsky

Ex-Staff Engineer at Meta. Building the technical foundation to make AI accessible to everyone.

Sergey Vysotsky

Reviewed by

Sergey Vysotsky

Ex-Chief Editor / PM at Mosaic. Passionate about making AI accessible and affordable for everyone.

Ready to experience the power of AI?

Access 200+ AI models, custom agents, and powerful tools - all in one subscription.