お知らせ • Jul 09
Kuaishou Unveils Comprehensive Ai Models, Reshaping Content and Commercial Ecosystem with AI
Kuaishou Technology made its first collective debut of AI models at a forum hosted on July 6th, titled "New AI New Applications New Ecosystem", as part of the 2024 World Artificial Intelligence Conference. During the Forum, Kuaishou showcased its comprehensive AI model matrix, including advanced functionalities for its video generation model "Kling" and image generation model "Kolors", among others. At the Forum, the third upgrade of Kling was released following the release of the image-to-video and video extension functions within the past month. Kling is now accessible via web portal. Together with the high-definition version of Kling, the new features unveiled at WAIC include start and end frame control and shot control capabilities. Additionally, the duration for single text-to-video generation for creators has been extended to 10 seconds. Kolors has been officially open-sourced to foster industry vitality and build a more prosperous text-to-image model community ecosystem. Mr. Gai outlined the Company's AI model matrix, which includes the KwaiYii (??) large language model, recommendation large model, and visual generation model as key components. These models span content creation, understanding, recommendation and other aspects, playing a crucial role in enhancing Kuaishou's commercial ecosystem. Notably, the recommendation model, SIM (Search-based Interest Model), with its scale of 10 trillion parameters, is one of the world's leading recommendation systems. Its next-generation architecture, ACT (Action Transformer), is expected to add hundreds of millions of minutes of daily user time spent on the Kuaishou App, significantly enhancing user engagement and activity. Drawing on the KwaiYii large model, Kuaishou has developed video script generation, real-time live streaming script generation, and advertising lead customer service, all integrated with digital human technology. These advancements help advertisers produce high-quality video and live streaming content affordably, thereby improving lead conversion efficiency. In June 2024, Kuaishou's peak daily spending from clients utilizing AIGC marketing materials exceeded RMB20 million, showcasing the enormous commercial potential of large models. Following the introduction of image-to-video and video extension functions, Kling has embraced its third major upgrade within a month. The web version is now officially online. During the Forum, Kuaishou announced a significant upgrade to Kling's foundational AI model, introducing enhanced high-definition quality as well as new editing capabilities like start and end frame control and shot control. Additionally, the maximum duration for single text-to-video creations has been extended to 10 seconds, marking the longest duration available to ordinary users in the industry for the present. Kling, the world's first video generation large model truly available to ordinary users, launched its text-to-video function on June 6th. At the Conference on Computer Vision and Pattern Recognition, it unveiled additional new features including image-to-video and video extension capabilities, enabling the creation of videos up to approximately three minutes in length. Based on real-world physical laws, the videos produced by Kling exhibit cinematic quality and dynamic effects, simulating lifelike physical movements with large motion and surpassing the constraints of traditional video generation technologies. This breakthrough has not only garnered praise locally but has also sparked considerable international attention, heightening global interest in China's advancements in AI technology. To date, over 500,000 users have applied for access to Kling's beta test, with the number of generated videos reaching 7 million. Popular creations such as "Old Photo Revival" have gone viral due to their emotional impact. Kuaishou will continue to focus on improving the model's foundational quality, enhancing video clarity and introducing more innovative features to meet diverse user needs. Mr. Wan Pengfei, head of Kuaishou's Visual Generation and Interaction Center, stated that the latest release of Kling brings significant enhancements in seven areas: motion generation, generation duration, adherence to physical laws, video quality, command response, image-to-video conversion and video control. These upgrades enable the creation of clearer and more manageable videos of 10 seconds or longer. Notably, the trailer for China's first original AIGC fantasy short play, "Legendary Mirrors of Mountains and Seas: Splitting Waves," premiered during the Forum, with Kling providing extensive technical support for the short play. The rapid advancement of AIGC technology has infused fresh vitality into the short play industry, significantly boosting the efficiency of short play production, creation, and operation. Furthermore, to inspire AI enthusiasts, Kuaishou launched the inaugural Kling x KuaiYing video creation contest "A Surge of Inspiration" at the Forum. This contest, in collaboration with six top institutions, boasts a prize pool exceeding RMB 300,000. Additionally, the contest launched the "Kling x Astral Short Plays" creator incubation program, inviting winners from each category to join a creator support program. This program offers notable visibility, cash rewards, and opportunities for direct engagement with industry professionals. In the field of image generation large models, Kuaishou's Kolors is at the industry forefront, boasting several core advantages including advanced semantic understanding, high-quality photographic visuals, and multi-condition controllable stylized generation capabilities. In the evaluation conducted by China's authoritative organization, the Beijing Academy of Artificial Intelligence Institute, Kolors scored 75.23, ranking second globally in the text-to-image model area. Kolors integrates Kuaishou's extensive expertise in large language models, trained on billions of Chinese-language data points, making it the most proficient Chinese text-to-image model available. Its overall performance outshines both open-source models like SDXL/SD3 and closed-source models like Midjourney, setting a new benchmark for image generation in Chinese contexts. During the Forum, Kuaishou announced that Kolors would be officially open-sourced, aiming to energize the industry and foster a more prosperous community ecosystem for text-to-image models. Additionally, the China Computer Federation (CCF) announced a collaboration with Kuaishou to establish the "CCF-Kuaishou Large Model Explorer Fund," with the fund application channel opening concurrently. This fund plans to launch 12 research projects by the end of 2024, with topics eligible for up to RMB300,000 in support. The fund aims to address the rapid development of AI technology and the industry's urgent demand for cutting-edge technology, focusing on the key technical research and development of the next generation of large models.
