Kakao (035720.KS) has developed a technology that reduces the computing costs of generating video with artificial intelligence and will present it at one of the world's three largest computer vision conferences.
Kakao said on the 10th that it will unveil the new technology, which addresses the massive computing loads and high costs involved in AI video generation, at the European Conference on Computer Vision (ECCV) 2026, running in Sweden through the 12th of this month. Lee Yeon-kyung, a research engineer at Kakao who led the work, will present the paper.
As research has expanded into high-resolution and long-form video generation, cutting computing costs has emerged as a central challenge for the industry. Existing video compression technology compressed data at a fixed ratio regardless of content. Even in scenes with a still background or little movement, it generated unnecessarily large numbers of tokens and consumed computing resources.
KATok, the adaptive tokenization technology Kakao developed, has the AI itself judge and adjust how much computation is needed based on the spatial and temporal complexity of the video. It assigns more tokens to scenes with heavy motion and dense information, while removing redundant tokens in static or simple scenes.
KATok reconstructs high-quality video using far fewer tokens than existing methods. In testing, the training speed of the video generation model improved about 6.9 times and video generation speed about 3.2 times compared with existing approaches. Kakao plans to further refine KATok and broaden its application to higher-resolution and longer-form video generation.
"The more computing loads surge, as with high-resolution and long-form video, the greater the computation savings that KATok's adaptive tokenization technology delivers," said Choi Dong-jin, model performance leader for applied AI at Kakao. "We will verify its general applicability with a range of AI video generation models and develop it into a core foundational technology for large-scale video generation that reduces computation while maintaining image quality."
