The researchers at the S Laboratory of Nanyang Technological University have open-sourced an innovative text-to-video model named FreeInit. They identified that video diffusion models often suffer from insufficient temporal consistency and unnatural dynamic effects when generating videos. To address this issue, they proposed a method to improve temporal consistency by reinitializing noise. They have made the code and paper of FreeInit available for researchers to reference and use. Through multiple iterative repetitions, FreeInit can further accumulate and enhance the quality of low-frequency information, gradually bridging the gap between training and inference initialization, thereby improving the quality and temporal consistency of the generated videos.