Seonglae ChoFlow Matching + Diffusion Transformer
Flow Matching connects noise→data via a probability flow (ODE), making it easy to directly learn a velocity field between frames for smooth and physically stable motion.
The ODE can be represented by a velocity field. The distribution of samples generated by this ODE satisfies the continuity equation (conservation of mass). In other words, when samples are pushed through the ODE, the resulting distribution shifts accordingly. From this perspective, the term "probability flow" emerges: probability mass flows according to the velocity field.
When combined with DiT, the training objective reduces to a simple MSE (velocity prediction), which provides good convergence stability and scales well even to billion-parameter models. It has lower hyperparameter sensitivity compared to traditional diffusion (SDE). The velocity-based objective is advantageous for reducing high-frequency artifacts like foot sliding and jitter. These benefits accumulate across pre-training→high-quality fine-tuning→RL stages. Since Flow Matching represents generation as ODE integration, it's easy to attach online RL like Flow-GRPO to directly optimize physical constraints (foot sliding, root drift) and semantic coherence as rewards. ODE solvers (e.g., Euler) produce stable results even with few steps, making it practical for long-sequence (several seconds) motion generation.
tencent/HY-Motion-1.0 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
https://huggingface.co/tencent/HY-Motion-1.0
