MotionRFT: Unified Reinforcement Fine-Tuning for Text-to-Motion Generation
MotionRFT introduces a framework to improve text-to-motion generation by combining a unified reward model with an efficient fine-tuning method called EasyTune. MotionReward creates a shared semantic space for different motion types, while EasyTune optimizes the model step-by-step during training instead of the whole trajectory. This results in better alignment with text, human preference, and realism across various motion styles.