Real models, in this tab
Each demo downloads an open model from Hugging Face and runs it right here with LiteRT.js — WebGPU when available, WASM otherwise. Nothing you load is uploaded.
Photo → 3D
One photo becomes an orbitable 3D point cloud. MoGe-2 monocular geometry, ~80 ms per photo on an M4 Max (WebGPU).
Open →Text → speech
Type a sentence and a local voice speaks it. Matcha-TTS flow matching + HiFi-GAN vocoder, 22 kHz.
Open →Image → text
A photo or screenshot becomes lines of text you can select and copy. PP-OCRv5 text detection (WebGPU) and recognition (WASM).
Open →Photo + words → mask
Type what to find in a photo — “a dog”, “the sky” — and get its mask, with no fixed list of classes. CLIPSeg, ~10 ms per new prompt on an M4 Max (WebGPU); a one-time ~280 MB download.
Open →Click to segment
Click anything in a photo and get its mask. SAM 2.1 Hiera-Tiny: ~60 ms per photo, ~4 ms per click on an M4 Max (WebGPU).
Open →Photo → objects
Boxes and labels on 80 kinds of everyday objects, in a photo or live from the camera. RF-DETR-Nano transformer detector, ~0.17 s per photo on an M4 Max (backbone on WebGPU, decoder on WASM).
Open →