Kunyang Li
Third-year Ph.D. Student in Computer Vision |
University of Central Florida, IAI | CRCV & Excel-Lab
Scroll Down
Third-year Ph.D. Student in Computer Vision |
University of Central Florida, IAI | CRCV & Excel-Lab
Scroll Down
I am Kunyang Li, a Ph.D. student at the
University of Central Florida (UCF)
,
jointly affiliated with the Institute of Artificial Intelligence (IAI) and the
Center for Research in Computer Vision (CRCV), where I am advised by
Dr. Mubarak Shah and
Dr. Yuzhang Shang.
My research interests lie in efficient video generation, video dataset distillation, personalized video generation, and video continual learning, with a particular focus on improving training efficiency and long-horizon generation in autoregressive and diffusion-based video models.
Previously, I received my B.S. degree from the
University of Electronic Science and Technology of China (UESTC)
and my M.S. degree from
Nanyang Technological University (NTU)
.
The video begins with a wide shot of a blue fishing boat gently navigating through the ocean, accompanied by a larger cruise ship in the background. The ocean is calm and clear, with a few small waves rolling gently. The fishing boat is moving slowly, following the trajectory of the cruise ship. The camera captures the details of the boat's design, including its sleek lines and the bright blue color. The surrounding environment is filled with the sounds of waves crashing against the boat and the gentle hum of the engines. The scene conveys a peaceful and serene atmosphere typical of a relaxing day at sea.
A training-free KV-cache management method for unified autoregressive video generation which dynamically compacts the KV cache through three coordinated mechanisms: condition anchoring, cross-frame decay modeling, and spatially preserving position embedding.
We propose GVD: Guiding Video Diffusion, the first diffusion-based video distillation method. Achieves 78.29% of original dataset's performance using only 1.98% of frames in MiniUCF, and 73.83% with just 3.30% of frames in HMDB51.
Learn More
We introduce PackCache, a training-free KV-cache management method which dynamically compacts the KV cache. Achieves 1.7–2.2× acceleration on 48-frame sequences, with 2.6× on A40 and 3.7× on H200 for the final frames.
Learn More
A comprehensive survey on PVG models (Sora 2, Veo 3.1). Covers VGF backbones, identity-preserving pipelines, open-domain & human-domain tasks, and real-world applications in e-commerce and gaming.
Learn MorearXiv 2025
A training-free KV-cache management method that dynamically compacts cache through condition anchoring, cross-frame decay modeling, and position embedding preservation. Achieves 1.7-2.2× acceleration on 48-frame sequences.
arXiv 2024
The first diffusion-based video distillation method. Achieves 78.29% of original dataset's performance using only 1.98% of frames in MiniUCF, and 73.83% with 3.30% of frames in HMDB51.
I'm always interested in discussing research collaborations, academic opportunities, or exciting projects in computer vision and AI. Feel free to reach out!
Center for Research in Computer Vision
University of Central Florida
By Appointment