School of Engineering Department of Computer Science and Engineering 106 Neural Rendering Supervisor: CHEN Qifeng / CSE Student: LU Zetian / CPEG Course: UROP 1100, Fall AnimateDiff has achieved remarkable results in generating videos with styles. However, the iterative sampling process across frames is computationally intense and leads to slow generation. In this UROP project, we studied the paper AnimateDiff, a text-to-video generation framework, then we worked on combining the AnimateDiff framework and the Latent Consistency Model to speed up the generation process. We found that this method can boost both the generation speed and the video quality. Neural Rendering Supervisor: CHEN Qifeng / CSE Student: YIP Tsz Hin / COGBM Course: UROP 1100, Spring Images taken by cameras often face physical limitations such as light source and distance, resulting in blurry or low dynamic range images. Advancements in machine learning have led to the development of various image super-resolution methods, evolving from deep learning models like SRCNN to more sophisticated generative models like GANs and diffusion models. The contribution to the project mainly falls into collecting and labelling public datasets suitable for training diffusion models. A total of 8000 images were processed, encompassing images with different angles of human face. This paves the way for the development of advanced image enhancement models for face super-resolution in the future. This report details the labelling methodology, challenges encountered, and related works done on the project. Trustworthy Machine Learning Supervisor: CHENG Minhao / CSE Student: BAI Yihan / COMP Course: UROP 2100, Fall Large language models (LLMs) have gained significant popularity due to their ability to produce text of exceptional quality. However, it is crucial to establish the origin of generated text to protect user copyrights, maintain security, and combat malicious activities. This report presents a comprehensive introduction to the utilization of watermarking methods for identifying a specific generator within a pool of given text. The proposed approach involves watermarking each generated token with user ID and previous context information. By leveraging this embedded watermark feature, the text from the specific generator can be accurately determined.
RkJQdWJsaXNoZXIy NDk5Njg=