This repository provides VPoser, a variational autoencoder that learns a low-dimensional latent representation (poZ) of human body pose from the AMASS dataset. It includes a PyTorch implementation of the SMPL body model (BodyModel), a VAE for pose encoding/decoding and sampling (VPoser), and an inverse kinematics engine (IK_Engine) that uses the learned prior to fit SMPL parameters to 2D/3D keypoints. The code is organized as a Python package with data preparation scripts, training pipelines using PyTorch Lightning, evaluation routines, and Jupyter tutorials. While primarily a research codebase for human pose modeling, it also serves as a reusable library for SMPL-based body representation and optimization. The repository is well-structured with modern dependency management (uv, pyproject.toml), CI testing, and clear separation of concerns, though it requires external downloads for the AMASS dataset and pretrained model weights.