brain.js is significantly simpler to deploy basic projects than TensorFlow.js

For object detection, we made a way to do it from scratch in a couple of hours (for free if your project is open source[1]): https://docs.roboflow.com/inference/web-browser

[1] https://blog.roboflow.com/computer-vision-datasets-and-apis/

I was checking it out, but it seems limited in terms of the type of models you can implement. You don't have CNNs for example.

CNNs are possible, but require some hacky code. Tensorflow.js is far better in this regard.

I used their facial landmark model, took some messing around (import/dep issues) but worked in my react project.

My use case was just automatic picture taking when a face was detected in a video stream.

also ONNX.js, which is great for interop: https://github.com/microsoft/onnxjs

Does it support GPU?