The pelvis bug should be fixed, it was introduced by the time consistency loss, which is anyway not needed as the vert_loss constrain well enough the SKEL fit, and bad global orientation initialization.
Here are some current fit examples for different SMPL pose datasets. They were generated with the current version of examples/align_to_SMPL_seq.py
CMU 05_04
CMU_05_04.mp4
MOYO Feathered_Peacock_Pose
moyo_feathered_peacock.mp4
DFaust 50002_chicken_wings
Dfaust50002_chicken_wings.mp4
MPI_Limits/03099/op5_poses