I have used both (albeit 2 years ago, and things change really fast). At the time, Candle didn't have 2d conv backprop with strides properly implemented. And getting Burn running libtch backend was just a lot simpler.

I did use candle for wasm based inference for teaching purposes - that was reasonably painless and pretty nice.