Inference Parameters

Overview

The most common inference arguments are:

• text_prompt: The description of the desired motion.
• num_samples: The number of samples to generate.
• packed_motion The input motion to be used as a starting point for the inference.

Instead of packed_motion, you can also infer from a local BVH file using:

• bvh_file: The path to the local BVH file to be used as input motion.
• edit_mode The masking mode applied onto the input motion.

Further options that might be interesting to adjust are:

• unpack_mode: How to unpack the motion and fill the missing values, can be stepped or linearly interpolated.
• unpack_randomness: The randomness applied during unpacking.
• editable_features: The features (joint positions, rotations and velocities) that are extracted from the input motion.

Finally, there are some experimental options:

• jacobian_ik: Whether to use Jacobian IK instead of Basic IK.
• foot_ik: Whether to use foot IK.

For an extended list of available parameters see the API Reference, specifically the InferenceArgs class.