Latent Pose

A data-driven IK demo in Blender
Pose sampled from a VAE(Vposer) trained with CMU mocap data.

My blog on this: Data-Driven IK In Character| 角色中的数据驱动反向动力学

Requirement

• Blender 2.8+
• python 3.6
  • pytorch 1.4.0
  • numpy 1.18
  • flask 1.1.2
  • tqdm 4.43

How to use

1. Start server, run python server.py

2. install blender plugin
   in Blender-Edit-Preference-Addons-Install, select latent-ik/latent-ik.py to install it.
   after installment, enable Latent-IK plugin

3. Open latent-ik.blend
   Next you can move controllers, use short-cut Ctrl+P to update pose
   Also you can select the Armature and enable or disable some controller
   

How it works

A trained VPoser on CMU mocap data

How to train your own vposer

1. download some bvh data into data/raw folder.
2. open data/raw/extract_skeleton.blend, import any bvh, run script data/raw/extract_skeleont.py to extract skeleton definition
3. run python -m data.extract_bvh_animation to extract bvh animation to joint's matrix into data/extracted
4. run python -m data.prepate_training to convert bvh animation to axis-angle form for training.

Reference

• VPoser
  https://github.com/nghorbani/human_body_prior
• SMPL-torch
  https://github.com/CalciferZh/SMPL