Upload an image to a vision-capable language model and get a short, plain description you can use as alt text. Tell it where the image will appear and what the page is about, then edit the result so it says what a person using a screen reader actually needs to know.
Alt text is the sentence that stands in for a picture when someone cannot see it: a blind reader using a screen reader, a slow connection that never loads the image, a search engine trying to understand your page. Most sites have hundreds of images and almost no alt text, because writing it by hand is dull. A vision language model can draft it in seconds, and you spend your time on the part that matters, which is checking that the description is true and useful.
Nobody skips alt text because they do not care. They skip it because a blog with three hundred images, a store with a few thousand product photos or a media library from five years of campaigns makes the job feel endless. A language model that can look at an image turns a blank field into a first draft, and editing a draft is far faster than staring at an empty box.
Picasso IA puts several vision-capable language models in the same toolkit, so you can try the same image on more than one and compare the wording. Gemini 3 Flash accepts up to ten images in a single request, which suits a batch of related photos from one article. Claude Sonnet 4.6 and GPT-4o each take an image together with a written instruction. The toolkit is where they all run, and the full catalog lists the other language models if you want to test more. New accounts get free credits, enough to run a real page through before you decide anything.
The honest framing is that AI is a drafting assistant here. It sees pixels, not your intent. It can tell that a photo shows two people at a whiteboard, but not that one of them is your new head of product and the point of the photo is the announcement. That context has to come from you, and the next section is about what to put in it.
Good alt text answers one question: what would someone miss if this image were gone? It is not a caption, not a keyword list and not a full inventory of every object in the frame. A reader on a screen reader hears it inline, in the middle of your paragraph, so every extra word costs them time.
A few habits separate useful descriptions from filler. Lead with the subject and the point, then add only the detail that supports the page. Skip openings like image of or picture of, because the screen reader already announces that it is an image. Describe the content, not the file. If the image contains text that matters, such as a sign, a button label or a chart title, include that text, since the reader cannot get it any other way.
The same photo needs different words on different pages. A picture of a ceramic mug is a product shot on a shop page, where color, shape and size matter. On a blog post about morning routines it is atmosphere, and one short phrase is plenty. This is the reason a model that only sees the image produces mediocre alt text, and a model that is also told the page and the purpose produces something close to publishable.
Different images do different jobs, so they need different instructions. Use this as a starting point for what to ask the model for, and what to expect it to get wrong.
| Image type | What the alt text should carry | What to check in the draft |
|---|---|---|
| Product photo | Item, color, material, visible features | Details the photo cannot prove, like size or fabric |
| Editorial photo | The subject and what is happening | Names of people, places and events the model cannot know |
| Chart or graph | The takeaway, plus where the full data lives | Numbers and trends read from the axes |
| Screenshot | The screen and the action it shows | Any interface text, copied word for word |
| Logo or icon | The name, or the action for a button icon | Whether it needs a description at all |
| Decorative flourish | Nothing, use an empty alt attribute | That the model did not describe it anyway |
Complex images such as charts and infographics need more than a sentence. A short alt text with the takeaway, plus the data in a table or a paragraph near the image, serves everyone better than a long block crammed into the attribute.
This is the routine that ends with descriptions you can publish, not a pile of generic captions. It works for one image or for a dozen from the same page.
A reusable prompt does most of the quality control for you. A good one names the page, the purpose and the limits in the same message, for example: this image appears in a post about repotting houseplants, describe what is shown in one sentence under twenty words, plain language, no guesses about people, and include any visible text exactly. Keep that as a system instruction and every later image inherits the same standards.
A description you did not read is not alt text, it is a liability. Models describe confidently and sometimes wrongly, and a screen reader user cannot see the picture to catch the mistake. Read every line against the image before it goes live.
Most of the problems you will meet repeat, so here is a short checklist to review before you paste anything:
The same drafting works for the image descriptions many social platforms let you attach, see AI social media posts, and product photography covers the store images you are likely describing.
This is where the draft ends and a person has to take over. A model cannot know why an image is on the page, so it will happily describe a decorative photo that should have been skipped, or flatten a photo whose whole point is a detail it treated as background.
It also cannot recognize the people you care about. Describing a group photo as three people smiling in an office is accurate and nearly useless if the page is about the founders, and the model should not be asked to identify real people from their faces. Add names yourself when the page calls for them, and only for people who are fine with it.
Charts, maps and diagrams are the hardest case. Models can misread axis values or invert a trend, and a wrong number delivered with confidence is worse than a missing one, so check every figure against the source data. Images that contain a lot of small text, like infographics and screenshots of long documents, are better handled by providing the text itself on the page.
Finally, accessibility obligations are yours to own. If your site has legal requirements, an automated draft does not meet them by itself, and the standard you follow, such as WCAG, asks for descriptions that are accurate and meaningful, not merely present. Treat the model as the first pass and a human review as the requirement. For video, the matching job is captions, which automatic video subtitles covers.
Alt text is a short written description attached to an image in your page code. Screen readers speak it aloud for people who cannot see the image, browsers show it when the image fails to load, and search engines use it as one clue about what the image contains. It matters most for accessibility: without it, a blind visitor hears only the file name or silence where your image should be.
Yes, as a first draft. A vision-capable language model such as Gemini 3 Flash, Claude Sonnet 4.6 or GPT-4o on Picasso IA can look at an image and describe what it shows. The draft is usually decent on objects and scenes and weaker on anything that depends on context you did not provide, so you review and edit each one before publishing.
Start with a language model that accepts image input. Gemini 3 Flash handles several images in one request, which helps when you are describing a set of photos from one article. Claude Sonnet 4.6 and GPT-4o take a single image with an instruction. Running the same image through two of them and comparing is a quick way to find the wording you prefer.
As short as it can be while still carrying the point of the image. For most photos that means one sentence, a handful of words to a short line. Complex charts are the exception, and for those a brief alt text plus the full data or explanation in the page text works better than a long attribute.
It helps search engines understand an image and the page around it, and it is the text shown if the image does not load. It is not a ranking trick. Descriptions written for people, with relevant words used naturally, serve both readers and search. Stuffing keywords into the attribute tends to read badly aloud and gains nothing, so write the sentence a person would want to hear.
No. Purely decorative images, such as spacers, dividing flourishes and background textures, should have an empty alt attribute so screen readers skip them. Describing them only adds noise for the listener. The test is simple: if removing the image would not change what the page tells someone, it is decorative, and the model should be told to leave it out.
It should not, and you should not ask it to. Identifying real people from their faces is unreliable and raises privacy problems. Describe what is visible, such as a woman presenting at a whiteboard, and add names yourself only where the page calls for them and the people involved are comfortable with it. For your own team photos, you know who is who and the model does not.
Write a short alt text that gives the main takeaway, then put the underlying numbers or the full explanation in a table or a paragraph next to the image. You can ask a vision model to read the chart and draft both pieces, but check every number against your source data, because models can misread axes or invert a trend without any sign that they did.
Within limits. Gemini 3 Flash accepts several images in a single request, so you can ask for one description per image in a numbered list. For a large media library, work in batches of related images and keep the same prompt, then review each result.
Generation uses credits, and the cost depends on the model and the settings you choose, so any figure written here would go out of date. Text responses from language models are generally among the lighter jobs in the catalog compared with video. New accounts get free credits to start, and the pricing page is the place to see the current plans.
Pick one page that has images without descriptions, open the Picasso IA toolkit and see how far a first draft gets you before you start editing.
Choose your language
