Lead Author & Primary Maintainer. Designed a general retrieval-augmented framework that lets long video generation models retrieve relevant prior content instead of relying solely on the recent context window.

Developed the inference engine for infinite video generation. Upgraded the existing architecture by designing an online relative RoPE to support infinite-length outputs.

Co-first Author & Primary Maintainer. Architected and implemented a novel evolving memory pipeline, enabling training-free stable infinite video generation.

Lead Author & Primary Maintainer. Pioneered successful demonstration of competitive liver tumor segmentation relying exclusively on synthetic training data.