Kakao announced the open-source release of two self-developed vision encoders, the 'Kanana-2 Vision Encoder,' on Hugging Face.
Vision encoders serve as the 'eyes' of AI, helping it understand images. By extracting visual features such as objects, text, locations, and shapes from images and relaying them to a language model, the model can interpret this information alongside user queries to generate responses. This technology is used for tasks like reading menu items and prices or understanding document content and layout.
The Kanana vision encoders include a 'Medium' model with 426 million parameters and a 'Tiny' model with 98 million parameters. Kakao previewed the open-source release of the vision encoders during the keynote at the 'if(kakao)26' conference held on October 13.
Kakao evaluated the performance of the models across nine areas, including general visual question answering and Korean-English text recognition and document understanding, using a total of 42 benchmarks. According to the company, both models achieved the highest performance compared to vision encoders from major global companies such as Google, Meta, Apple, and Alibaba. Notably, the lightweight Tiny model demonstrated strong performance in Korean document understanding and image location detection, even when compared to models with over 400 million parameters.
Utilizing its self-developed 'Kanana-2' language model, Kakao has increased the learning efficiency of the vision encoders by approximately four times and enhanced their ability to extract detailed visual information.
Following the open-source release of three Kanana-2 language models in December last year, Kakao introduced four improved models in January. In July, the company also launched four lightweight language models capable of running on devices like smartphones, further expanding the availability of its AI models.
Kakao will release the two models under the Kanana open license, allowing for both research and commercial use. This aims to enhance accessibility for researchers and companies both domestically and internationally, supporting advancements in multimodal AI research.
Noh Byeong-seok, leader of Kakao's Unified Foundation Model Performance, stated, "As we enter the era of agentic AI, the ability to directly understand and manipulate PC or mobile screens is gaining attention, increasing the importance of vision encoders that accurately analyze high-resolution images. We will continue to enhance the competitiveness of our AI models and contribute to the development of the AI ecosystem by making related technologies available."
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.