Work Experiences

Research on generative models. During this period,

• A high-quality text-to-video generation framework LaVie is proposed.
• A general Transformer-based latent video diffusion model, referred to as Latte, is introduced.
• An image animation framework Cinemo based on motion diffusion models is introduced.

Research on model compression. A model compression tool has been developed to assist developers in rapidly deploying models to edge devices without compromising model accuracy, while simultaneously enhancing model inference speed. This tool has been extensively implemented across various businesses at Meituan. One paper was accepted by CVPR 2022 during this period.

Research on Image Dewatermarking Algorithm. An image dewatermarking algorithm was proposed based on attention mechanism and self-supervised learning. The service is now launched on Meituan App. Related work was accepted by ICPR 2020 and selected as an oral presentation.