How to use the Local AI Vision Lab
- Check “AI capabilities of this device”: WebGPU plus WASM SIMD gives the best speed.
- Pick a tool by task: fitness, documents, product photos, privacy or kids learning.
- The first run downloads that tool’s model (about 0.2–16 MB); later visits load it from the browser cache.
- Copy or download the result, and jump to the background remover, upscaler or image editor for fine touch-ups.
What does “local AI” mean?
Most online AI tools upload your image to a server, run a cloud model and send the answer back. Local AI downloads a compact model (from a few hundred KB to a few dozen MB) into the browser and runs it on your own computer or phone with WebAssembly or the GPU. Your photos stay private, there is no queue, and there are no usage limits. The trade-off is a one-time model download and slower runs on older devices.
Which models are used?
Body pose uses Google MediaPipe Pose Landmarker (lite, about 5.8 MB); faces use BlazeFace (about 0.2 MB); objects use EfficientDet-Lite0 (80 COCO classes); image classification uses EfficientNet-Lite0 (1,000 ImageNet classes); text recognition uses Tesseract LSTM models for English and Chinese; cutouts reuse the MagicTouch interactive segmenter already powering our background remover. All are open-source (Apache-2.0 and similar), hosted on this site and verified by SHA-256 before use.
Where does Chrome built-in AI (Gemini Nano) fit?
Recent desktop Chrome can download a built-in Gemini Nano model that web pages reach through the Prompt API. We treat it as an optional enhancement: in tools such as the alt text generator and the kids identifier, an extra button appears when your browser supports it. Without it, the cross-browser models above still do the job.
How do I get the best speed?
Use the latest Chrome or Edge on a computer with up-to-date graphics drivers, or the latest built-in browser on a phone. Camera tools work best in good light against a simple background. The first load is slower because the model is downloading; after that it is cached.
Frequently asked questions
Are my photos and camera frames really not uploaded?
Correct. All inference happens in the browser. The page only downloads model and program files from this site and never sends images, video frames or results to a server. Once a model is cached you can even go offline and keep working.
Do I need an account or to pay?
No. Every tool is free, needs no sign-in and has no usage limits.
Why does the first run download over 10 MB?
That is the AI model and its runtime. Your browser caches them, so later visits start almost instantly without using more data.
Does it work on phones?
Yes. Current iOS Safari and Android Chrome can run these models, though low-end phones load more slowly and real-time camera tracking may lag. A laptop or recent phone gives the smoothest experience.
How accurate are the results?
These are small general-purpose models meant for everyday reference. Rep counting and OCR are dependable with good light and angles; food calories and species identification are estimates or guesses, and each page explains its limits.
How does this relate to the existing background remover and upscaler?
They belong to the same local AI product line. This hub groups them with the new tools, and the Product Photo Studio chains cutout, sizing and compliance checks into one workflow.