Kuaishou–Tsinghua Joint Research Institute Showcases AI and Video Technology Collaboration at the Software Discipline Development Forum

On April 24, 2021, the Tsinghua University Software Discipline Development Forum and the 20th anniversary ceremony of the School of Software were held in the FIT building; the Kuaishou–Tsinghua Future Media Data Joint Research Institute co‑organized the event.

Dr. Zheng Wen, deputy director of the institute and head of Kuaishou Y‑Tech, spoke as the first undergraduate graduate representative, expressing gratitude to the school, extending blessings for the university’s 110th anniversary and the institute’s 20th anniversary, and sharing insights on research innovation, discipline construction, and talent cultivation achieved through industry‑academia collaboration.

Leveraging the institute’s platform, Kuaishou collaborates with more than twenty professors from Tsinghua’s Software School, Automation Department, Computer Science and Technology Department, and Electronic Engineering Department across numerous fields.

Kuaishou aims to jointly explore video‑centred AI technologies such as low‑light enhancement, real‑time denoising, motion‑object detection, mobile‑side real‑time computer‑vision (hand/gesture capture and 3‑D reconstruction), multimodal learning for video content, and large‑scale recommendation‑system hardware‑software co‑optimization (e.g., NVM‑based massive‑scale recommendation indexes and heterogeneous acceleration platforms). These research outcomes have been applied in Kuaishou’s main‑site app, live‑streaming, and other core short‑video industry chains, and have resulted in more than ten papers at top conferences (NeurIPS, CVPR, Supercomputing, ICDE, AAAI, IJCAI, ACM MM), over twenty national invention patents, and award‑winning technologies such as the 2019 CIE technical invention first prize and the 2020 Chinese AI Society Wu Wenjian AI technology progress first prize.

The partnership also nurtures talent, jointly training dozens of outstanding graduate students.

At the event, the institute presented an offline “Magic Table” interaction showcasing Kuaishou’s portrait‑style transfer effects. The portrait‑style transfer, powered by AI algorithms, converts input portrait images into target‑style images with high fidelity in overall style and fine‑grained details. A self‑developed Mixed CycleGAN architecture, built on the traditional CycleGAN model, was trained using paired manually retouched data and high‑quality unpaired data, enabling rapid convergence to target‑style results. By incorporating facial key‑points and semantic segmentation, the system precisely enhances local details while preserving global consistency.

During Tsinghua’s 110th‑anniversary celebrations, Kuaishou was authorized as one of the media broadcasters. Using its underlying live‑streaming infrastructure and extensive VR‑panorama experience, Kuaishou delivered a unique mobile‑intelligent shooting and full‑chain VR live‑streaming solution, covering video capture, stitching, encoding, pushing, transmission, transcoding, storage, distribution, decoding, and playback. Viewers could switch among multiple live rooms to experience 360° campus panoramas and conventional streams, with a Tsinghua professor hosting a “cloud tour” that started from the university’s main gate.