Episode Details

Recent progress in video generation still struggles with issues like motion instability and prompt alignment. To address this, the study explores incorporating human preferences into advanced flow-based video generation models. The authors introduce a large, new dataset of human-annotated video preferences across visual quality, motion quality, and text alignment. They also develop a multi-dimensional reward model to quantify these preferences and propose three alignment algorithms for flow-based models, demonstrating that a modified Direct Preference Optimization method yields the most effective results in aligning video generation with human expectations.