School of Engineering Department of Computer Science and Engineering 116 Generative AI Supervisor: CHEN Qifeng / CSE Student: RUAN Junhong / COSC Course: UROP 2100, Fall UROP 3100, Spring In the work of robust 3D reflection removal via 3DGS, we try to implement Vanilla 3DGS to get a reflectionfree depth map for every frame of a video, and using a depth-based warping process to extract a transmission-only pseudo ground truth for every single frame as a supervisory in the training of our model. Since it is difficult to obtain a real ground truth, we try to construct virtual scenes with different settings and transmissions using rendering engine Unity. Generative AI Supervisor: CHEN Qifeng / CSE Student: SAKHMOLDIN Mukhammadarif / DSCT Course: UROP 1100, Spring Recent advancements in Large Language Models have led to the rise of Multi-modal Large Language Models that can process and generate content across text, images, and audio. However, the computational resources required for these models present significant challenges for widespread deployment. During UROP we explored knowledge distillation techniques as a solution, focusing on transferring capabilities from large teacher models to more efficient student models. We examine frameworks like LLaVA-MoD, which employ KL Divergence distillation methods and Mixture-of-Experts architectures to create compact yet powerful MLLMs. Our experimental work includes testing models, conducting adaptor pretraining, and performing evaluations on benchmarks. Generative AI Supervisor: CHEN Qifeng / CSE Student: SUN Mengxi / DSCT Course: UROP 2100, Fall In this report, I summarize what I have done and learned from UROP 2100 in this fall term. During this term, I continued participating in the same project group in UROP 1100. The focus of this project is to design a visual language model to process images and give suggestions on how to adjust the camera parameters tailored to users’ needs. In this term, I collected certain amounts of required pictures for the project, studied on bokeh effect and tried to modify PyNET code. I also joined several group meetings, from which I started to understand their work. In the end, I feel grateful of having this opportunity to participate in the research process and explore a new era.
RkJQdWJsaXNoZXIy NDk5Njg=