Bilibili's Self-Developed Video Super-Resolution Algorithm: Background, Optimization Directions, and Implementation Details

Bilibili has developed a self‑supervised video super‑resolution algorithm to enhance low‑resolution content to 4K for a clearer viewing experience.

The article first outlines the gap between 1080p and 4K video, showing that traditional interpolation methods (nearest‑neighbor, bilinear, bicubic) can upsample resolution but fail to recover lost texture details.

It then explains why deep‑learning‑based super‑resolution is needed, demonstrating via visual comparisons that it reduces artifacts and restores sharper edges and finer textures.

Through DCT‑based frequency analysis of massive video libraries, the authors categorize content into anime, games, and real‑world footage, assigning specific optimization directions: sharpening lines and removing noise for anime; balancing line sharpness, noise removal, and texture strengthening for games; and enhancing fine lines and complex textures for real‑world videos.

The self‑designed algorithm consists of three parallel degradation branches (texture‑enhancing, line‑recovering, noise‑removing) whose outputs are weighted according to video type, a model built from stacked REPB blocks (combining Conv‑x and ESA modules), and a composite loss function that blends pixel, line, and texture terms with content‑dependent weights.

Finally, the algorithm is deployed in B站’s online point‑to‑play pipeline, delivering measurable quality gains, and the team outlines future work on broader coverage, subjective improvement, and deployment flexibility across live, VOD, and client scenarios.