This paper presents an inverse kinematics (IK) system that learns from a probability distribution of human poses to generate realistic and style-specific poses in real-time. The system is trained on a set of constraints and can produce the most likely pose satisfying those constraints. The probability distribution is represented using a Scaled Gaussian Process Latent Variable Model (SGPLVM), which automatically learns the parameters without manual tuning. The SGPLVM captures the likelihood of poses, allowing the system to generate any pose but preferring those similar to the training data. The paper also introduces a novel method for interpolating between different styles of poses. The system is demonstrated in various applications, including interactive character posing, trajectory keyframing, real-time motion capture with missing markers, and pose estimation from 2D images. The main limitation is the need for suitable training data, and the system does not explicitly model dynamics or constraints from original motion capture data. However, it produces more natural poses compared to existing approaches, even with generic training data.This paper presents an inverse kinematics (IK) system that learns from a probability distribution of human poses to generate realistic and style-specific poses in real-time. The system is trained on a set of constraints and can produce the most likely pose satisfying those constraints. The probability distribution is represented using a Scaled Gaussian Process Latent Variable Model (SGPLVM), which automatically learns the parameters without manual tuning. The SGPLVM captures the likelihood of poses, allowing the system to generate any pose but preferring those similar to the training data. The paper also introduces a novel method for interpolating between different styles of poses. The system is demonstrated in various applications, including interactive character posing, trajectory keyframing, real-time motion capture with missing markers, and pose estimation from 2D images. The main limitation is the need for suitable training data, and the system does not explicitly model dynamics or constraints from original motion capture data. However, it produces more natural poses compared to existing approaches, even with generic training data.