tract in your browser

Neural networks running in your browser, on your GPU through WebGPU or on your CPU, with tract, compiled to WebAssembly. Nothing leaves your machine.

Stable Diffusion 1.5

Text to image, 512 × 512, with a prompt, a negative prompt and a choice of sampler.

Downloads 2.1 GB of fp16 weights from Hugging Face on the first visit. Needs WebGPU with shader-f16.

Person segmentation on your camera, with the background blurred or replaced by a colour, live.

Uses your camera. Downloads a 0.5 MB or 26 MB model from Hugging Face. Needs WebGPU.

Streaming transcription of your microphone with NVIDIA's Nemotron 3.5 ASR, in dozens of languages.

Uses your microphone. Downloads 2.6 GB of weights from Hugging Face on the first visit. Needs WebGPU.

Wake word detection on your microphone, streaming, from raw samples to a score every 10 ms.

Uses your microphone. Runs on the CPU in WebAssembly, no WebGPU needed. 1.2 MB model.

Find the text in a sample page, a picture you choose, or a camera snapshot.

PP-OCRv6 Tiny, Small, or Medium from Hugging Face. Runs on tract, on the CPU or on WebGPU.

Source code of the demos, in tract's examples.