Best framework for DNN module

The DNN module in OpenCV supports networks exported from several frameworks, like Caffee, Tensorflow, Torch, ONNX and Darknet. However the support is still partial, some data types, layer types and parameters are not implemented, leading the loader function to abort.
This is quite important to me, as OpenCV is one of the few libraries providing C++ support for DNN inference.

What is your experience with the current support of frameworks? Which works best, especially for state of the art networks?
And which one is the simplest to use? (like Tensorflow networks require to be “freezed”, which was quite complicated last time I tried for TF2).

Observation: I found that Darknet works the best, but it already has a C++ interface and unfortunately it is not well maintained anymore. So it supports well Yolo and some classification networks, but that’s mostly it.