This implements a new adaptive integrator that switches to forward sampling after the model loss is better than the loss obtain on the same sample but using the identity as a flow.
Caveat I use the standard deviation of the identity model loss to estimate whether the model is significantly better. Accessing the standard deviation of the model loss is quite painful and would require changing a lot of the internals - better leave that alone for now.
This implements a new adaptive integrator that switches to forward sampling after the model loss is better than the loss obtain on the same sample but using the identity as a flow.
Caveat I use the standard deviation of the identity model loss to estimate whether the model is significantly better. Accessing the standard deviation of the model loss is quite painful and would require changing a lot of the internals - better leave that alone for now.