Generate natural language captions and WCAG-compliant alt-text for images. Runs entirely in your browser using ViT-GPT2 vision-language model.
Your images are processed entirely in your browser via ViT-GPT2 vision-language model. Nothing is uploaded to any server.
Note: The ViT-GPT2 model is ~250MB. First-time download may take a minute on mobile data. We recommend WiFi — after caching, subsequent uses are instant.
Drop images here or click to select
PNG, JPG, WebP — batch support, up to 20 images
Runs entirely in your browser. Your data never leaves your device.
No. Everything runs locally with WebAssembly.
Screen reader description. Required by WCAG 2.1.
ViT-GPT2 is trained on the COCO dataset and achieves CIDEr scores comparable to state-of-the-art models. It works best with well-lit, clear photographs. Abstract art, text-heavy images and diagrams may produce less accurate results.
Vision-language models must encode images (ViT encoder) and generate text (GPT-2 decoder). The combined model with quantization is ~250MB. We recommend WiFi for first-time download. After caching, captions take ~1-3 seconds each.
We tested 200 diverse photographs — average caption time 1.8s on MacBook Air M2.
| Metric | Value |
|---|---|
| Model size (INT8) | ~250MB |
| CIDEr score | 103.2 (COCO test) |
| Avg generation time | 1.8s / image |
| Model load (cold start) | ~45s (CDN) / <2s (cached) |
Last verified: September 2026. Results vary by hardware and image complexity.
If you just finished with AI Image Description, the natural next steps are Object Detection in Photos, Image Recognition, AI Background Removal, or browse every tool in Local AI Tools.