We are delighted to announce that 1 paper from VE Lab (Prof. Youngho Chai) have been accepted to IEEE Access.

Title: 

Anatomical Control in Text-to-Motion Diffusion Using Joint Condition Vectors

Authors:

Soungsill Park, Jeongyeon Lee, Kyungmin Kim, Youngho Chai

Abstract:

We present a text-to-motion diffusion framework that integrates explicit anatomical control using a Joint Condition Vector (JCV). The JCV specifies the minimum, maximum, and start angles for six lower-body channels—left/right hip abduction/flexion and knee flexion—resulting in an 18-dimensional control vector. We condition the leg encoders with the JCV through Feature-wise Linear Modulation (FiLM) and apply a conditional loss to Euler angles reconstructed from the predicted motions. To improve rotational validity and stability, we replace inverse kinematics (IK)-based recovery with an AMASS-aligned SMPL representation and adopt a numerically safe 6D-to-Euler conversion to reduce discontinuities. We construct HumanML3D-SMPL and apply targeted lower-body motion augmentation to enhance diversity while preserving anatomical validity. On the HumanML3D-SMPL test set, our method improves distributional fidelity in the T2M evaluation space (FID 0.40 vs. 0.47), while maintaining competitive performance in the TMR latent space. These results indicate that JCV conditioning improves distributional fidelity while maintaining competitive generation quality across multiple evaluation spaces. Qualitatively, the model produces anatomically interpretable variations (e.g., abduction vs. flexion) from the same text prompt and composes unseen primitives, such as shuffle-dance steps, through segment-level generation and overlap-and-blend sampling. These results demonstrate that explicit joint-level conditioning can be achieved without compromising diffusion-based generation quality. The proposed framework bridges semantic intent and biomechanical precision, supporting controllable and anatomically accurate motion synthesis.