I'm building a tool that requires the observed keyframes to be unchanged. What's the best way to do this?
I trained the model with zero_keyframe_loss = True and velocity loss to mitigate jitter. But there is always a pop between the generated and ground truth at the observed keyframes.