Pick one or more images and get a draft description for each - a starting point for the alt text on your site. Everything runs on your device; the images are never uploaded.
Sample alt text this page can draft for an image (nothing is uploaded, and this preview needs no model download to read):
Caption the on-device model produced for this exact image: a woman in a blue shirt and a blue and white flag
As a copyable attribute: alt="a woman in a blue shirt and a blue and white flag"
The caption above was produced ahead of time by the same model this page runs, so you can see the output shape before loading anything. It is a literal machine guess - the photo shows astronaut Sally Ride (NASA, public domain), which the model cannot know; that is exactly why every caption here needs your review.
Alt Text Generator - Describe Images On-Device
This page turns images into draft alt text without sending them anywhere. Pick up to 20 images, and each one gets a short machine-written description plus a ready-to-paste alt="..." attribute. Nothing you select ever leaves your device.
How the describing works
The describing happens in your own browser, on one of three tiers. When your Chrome has built-in on-device AI, captions need no download from this page at all - the built-in engine is feature-detected, never assumed. Otherwise you load a caption model once with a single click: the honest size (about 250 MB, one time, cached by your browser) is stated on the button before anything starts downloading. On every tier the only network download is model weights, or nothing at all - your images are never uploaded.
Key features
- Describe up to 20 images in one batch; each gets its own editable caption row with a copyable alt="..." attribute, plus a copy-all button at the end that collects every finished attribute in one paste.
- Reads the scene, not the words - it describes what the picture looks like rather than extracting printed text.
- Every draft lands in an editable box so you can correct it, cut it to the one fact a reader needs, then copy the attribute.
Read every caption before you publish
The honest part first: a machine caption is a literal guess, not finished alt text. The model describes what pixels look like ("a woman in a blue shirt"), and it can get people, colors and context wrong. That is why every caption is editable - read it, correct it, then copy. If an image is purely decorative, the right answer is an empty alt attribute, and the per-image hints on this page say so. A sample image with its real precomputed caption is shown on the page before anything loads, so you can see the quality first.
No download needed to start
You do not need any download to get value here. With no model loaded, the same batch view becomes a writing workbench: each image shows its thumbnail, dimensions and a short guidance line matched to what the file looks like (chart, logo, screenshot or banner), with the same editable field and copyable output. Load the on-device model only when you want machine drafts for a large batch.
Built for batches
Batch is where this saves real time. A gallery page, a product listing or a documentation set often needs dozens of descriptions in one sitting; queue the images, let each one get a draft in a few seconds, and spend your attention on editing instead of typing every sentence from scratch.
Related tools
Two neighbouring jobs have their own pages. If you need the words printed inside an image - a sign, a receipt, a screenshot - that is optical character recognition, and the image to text (OCR) tool does it on-device the same way. And if a photo needs faces hidden before you publish it at all, the face blur tool handles that first, also without uploading anything.
Frequently Asked Questions
Are my images uploaded anywhere?
No. The images are read and described in your own browser on every path this page offers. The only network download is the caption model itself (about 250 MB, one time, kept by your browser), or nothing at all if your Chrome already has built-in on-device AI or you use the writing workbench.
Can I publish the generated captions as-is?
No - treat them as first drafts. The captions are literal scene descriptions from a small on-device model; it can misidentify people, colors and context, and it does not know names, brands or intent. Read each caption, fix what is wrong, and shorten it to what a reader actually needs.
Why is the download so large, and do I pay it every time?
The caption model is about 250 MB at its compressed size (an 87 MB image encoder plus a 159 MB text decoder). Your browser caches it after the first load, so later visits caption with no new download. The size is shown on the button before anything starts.
Does it read the text written inside an image?
No. This page describes the scene ("two cats laying on a bed"); it does not extract printed words. To pull the text out of a photo or screenshot, use the image to text (OCR) tool instead.
Can I caption many images at once?
Yes - pick up to 20 images in one go. They are processed one after another on your device, each gets its own editable caption and a copyable alt="..." snippet, and a copy-all button collects every attribute at the end.