Kuaishou engineers have introduced the Large Processing Model (LPM), the industry's first generative diffusion-based system designed for large-scale video restoration. Unlike traditional methods, LPM is capable of not just increasing resolution, but actually reconstructing lost details such as faces, textures, and text, while ensuring high temporal stability for long clips.
What Happened
Kuaishou has developed the Large Processing Model (LPM), which utilizes a Temporal-Pyramid Inference mechanism to ensure video sequence stability and eliminate flickering. The system is already integrated into the Kuaishou ecosystem, where it allows for a 20% reduction in video bitrate while maintaining visual quality, providing significant savings on traffic transmission infrastructure.
Context
Modern video platforms face the challenge of massive costs for data storage and transmission, as well as the need to process low-quality user-generated content. Traditional upscaling algorithms are often limited in their ability to restore truly lost details, creating a gap between old archives and modern quality standards.
Why It Matters for the Industry
For the industry, this signifies a shift from simple upscaling to full-scale generative post-processing of media streams. The technology demonstrates the possibility of creating complex ecosystems where the focus shifts from generating content "from scratch" to its intelligent optimization and scalable enhancement, which is critical for reducing infrastructure costs for large platforms.
Why It Matters for Users
Users will experience a qualitative leap in content consumption: even old or poorly recorded social media videos will look like modern high-quality clips. Additionally, new generative videos will become more stable, detailed, and visually pleasing by eliminating compression artifacts.
What Remains Unknown / Limitations
There are concerns regarding the risk of model "hallucinations," where generative processes might distort original content, such as altering facial features or text. To combat this, LPM uses a specialized training method that preserves authentic parts of the video.
Sources
Author
Look at AI, Editorial Team