Data dyeing attaches one 512-dimensional text-aligned MotionCLIP latent to each collected frame. The public runtime loads only the encoder and CLIP text model needed for this step.
Place the released MotionCLIP checkpoint at:
Artifacts/checkpoints/motionclip/g1-model-xyz-clip/checkpoint_0100.pth.tar
After collection, run from the public repository root:
uv run python PredActor/collection/scripts/data_dyeing/data_dyeing.pyThe default configuration reads:
Artifacts/dataset/g1_152fk_raw/motion.zarr
Artifacts/motions/g1_152fk/*/motion.npz
and writes:
Artifacts/dataset/g1_152fk/motion_dyed_text_t0.01.zarr
Paths can be overridden with Hydra values such as
input.zarr_path=/path/to/motion.zarr and
output.zarr_path=/path/to/dyed.zarr. The released profile uses the original
motions for encoding, 60-frame windows, 30 Hz target sampling, xyz poses,
and text-alignment temperature 0.01.