A curated list of ML models that run fully in the browser. No server. No API key. No data leaves the device.
Powered by WebAssembly, WebGPU, and ONNX Runtime Web.
- background-removal-js - Remove image backgrounds in browser via WebGPU with WASM fallback. Based on RMBG-1.4
- WaifuXL - 4x image upscaling via tiled WASM processing. Based on Real-ESRGAN
- DeOldify-on-Browser - Colorize B&W photos client-side via ONNX Runtime Web. Based on DeOldify
- tfjs-models/face-detection - Face blur and anonymization via WebGL at 30fps+. Based on BlazeFace
- transformers.js - Depth map generation from a single image via WebGPU. Based on MiDaS
- yolov8-tfjs - Real-time object detection on webcam via WebGL. Based on YOLOv8
- tesseract.js - OCR text extraction via WASM SIMD, 100+ languages. Based on Tesseract C++
- rnnoise-wasm - Real-time noise suppression inside AudioWorklet via synchronous WASM. Based on RNNoise
- vad-web - Voice activity detection via ONNX Runtime Web, resamples to 16kHz. Based on Silero VAD
- whisper.cpp - Speech to text via WASM SIMD, Whisper tiny runs at ~75MB. Based on OpenAI Whisper
- mediapipe/tasks-audio - Sound classification across 521 classes via WASM. Based on YAMNet
- tfjs-models/pose-detection - Pose estimation tracking 17 keypoints at 50fps+ via WebGL. Based on MoveNet
- tfjs-models/hand-pose-detection - Hand tracking with 21 3D landmarks per hand via WebGL. Based on MediaPipe Hands
- tfjs-models/body-segmentation - Body segmentation for virtual backgrounds via WebGPU. Based on MediaPipe SelfieSegmentation
- kan - Eye strain and blink detection via Eye Aspect Ratio calculation. Based on MediaPipe Face Mesh
- fasttext.wasm.js - Language detection for 176 languages under 1ms, 900KB model. Based on fastText
- transformers.js - Sentiment analysis via DistilBERT with WebGPU acceleration. Based on distilbert-sst-2
- transformers.js - Text classification and spam detection via quantized BERT + WASM
The three core runtimes powering everything above:
- ONNX Runtime Web - Runs any ONNX model via WebGPU, WASM, or WebNN. Most versatile option.
- Transformers.js - Hugging Face models in the browser. Same API as Python transformers. 200+ architectures.
- TensorFlow.js - Mature ML framework with WebGL and WASM backends. Best for MediaPipe and MoveNet.
- MediaPipe - Google's optimized tasks for vision, audio, and body tracking. Production-ready.
- awesome-webnn - WebNN API resources and hardware acceleration
- awesome-webgpu - WebGPU ecosystem and compute shader tooling
- awesome-on-device-AI - On-device AI for mobile and edge devices
- awesome-webmcp - WebMCP standard for exposing browser models to AI agents
This list started as a personal reference while building browser-based ML tools at work. No server, no API key, no data leaving the device was the constraint. Finding the right repos was harder than it should be.
If you know a model that runs fully in the browser and belongs here, open a PR. One rule: working browser implementation required. Python repos and papers go elsewhere.
GitHub links only please. Your contributions are appreciated.
Author: Lucian Crisan