China Telecom's Artificial Intelligence Research Institute (TeleAI) has released a groundbreaking generative video compression technology - GVC (Generative Video Compression). This technology has increased the video data compression rate to an astonishing 0.02%, meaning that a 1GB video file can theoretically be transmitted with only about 200KB of data, and the video can still be viewed with clear image quality.

image.png

The core logic of this technology is called "Trading Computation for Bandwidth". Unlike traditional video encoding (such as H.265 or H.266), which relies on "moving pixels", GVC no longer transmits complete picture pixels, but instead sends "instructions on how to draw the picture". These small data packets are called "compressed Tokens", which contain semantic information of the scene (such as the structure of objects) and motion information (such as movement trends).

At the receiving end, a pre-trained generative model acts as a "painter". It generates a coherent and realistic video based on the received Token instructions, combined with its extensive knowledge of the world (such as the visual features of waves or footballs). This mode directly bypasses the problems of image breakdown and lag that traditional technologies often face in extremely low bandwidth scenarios.

According to the technical report published by TeleAI, GVC performs significantly better than traditional algorithms on authoritative datasets. At the same visual quality, traditional methods consume more than six times the bandwidth of GVC. Currently, the model can already achieve near real-time generation speed on consumer-grade graphics cards (such as RTX4090). This technology has the potential to solve the urgent need for high-definition video transmission in extreme network environments such as long-distance communication, emergency rescue, and deep space exploration in the future.

    Technical Report Address:

    https://www.arxiv.org/abs/2512.24300

Key Points:

  • 📉 Extreme Compression: The technology has compressed the video to 0.02%, allowing a 1GB video to be restored and presented at the receiving end with only 200KB of data.

  • 🧠 Logical Shift: It changes the traditional pixel transportation model, transmitting high-dimensional semantic Tokens and using generative AI at the terminal to "redraw" the video.

  • ⚓ Broad Application: Designed for extremely low bandwidth environments, it can be applied to satellite communications, long-distance voyages, and disaster site rescue in extreme signal scenarios.