“--predict_smpl” fails at init-pose loading, and the produced SMPL sequence is nearly static

[johnadams0913]

Hi, thanks for releasing CoMoVi! I'm trying to get the 3D SMPL output via --predict_smpl and hit two related problems.

1. inference.py writes the reference-frame SMPL via CameraHMR's process_image, which does np.save(frame_result). But get_smpl_init_pose (comovi/utils/utils.py:346-351) reads it with np.load(...) and then indexes smpl_npz_data["smpl"] and ["camera"], expecting a multi-frame .npz. So "python inference.py --predict_smpl" crashes before the pipeline runs.
2. After fixing the loader, the predicted SMPL motion is almost static. Across many reference images/prompts, the 80-frame sequence stays essentially at the initial pose.

example_output_triple.mp4

Could you help me how to figure these two issues?