Shulou(Shulou.com)11/24 Report--

Recently, Kuaishou began to test "Kuaishou AI comments" internally in short video comments, where users can type in a variety of creative words and generate massive style pictures with one click, making it easier and easier to interact with each other in the comment area. This is another AIGC ability for Kuaishou to land in a short video scene after "AI conversation".

It is reported that. Kuaishou "AI play Review" relies on the powerful image generation ability of Kuaishou's self-developed large model "Kolors". The Ketu large model can generate paintings with various styles, exquisite picture quality and creativity based on open text, so that users can easily and efficiently complete artistic creation. This is also the latest breakthrough and layout in the field of AIGC announced by Kuaishou AI team after the big language model "happy".

With the continuous breakthrough and innovation of AI technology, AIGC will become a "boundary-breaking" tool for content creators in the future, and it will enable more people who "have pictures in their minds and stories in their hearts" to create more easily and efficiently. It is understood that the three characteristics of the large model ensure an excellent image generation effect:

First of all, strong text understanding ability. Kuaishou AI has built billions of graphics and text training data from the open source community, internal construction and self-developed AI technology synthesis. These data cover the common 30 million Chinese entity concepts and have world knowledge. On this basis, we train and develop a powerful Chinese CLIP model, which not only understands our language, but also understands the images of the Chinese world. Secondly, Kuaishou AI uses the self-developed Chinese LLM, merging the picture and text features of CLIP as the text understanding module of CLIP, which not only realizes the understanding of Chinese characteristic concepts, but also solves the common problems in the field of literary graphics, such as complex concepts, attribute confusion and so on.

Secondly, a wealth of detail description. The Kuaishou AI research team changed the underlying formula and noise adding formula of the denoising algorithm; at the same time, it selected a number of high-detail, high-aesthetic high-quality data to focus on learning in the later stage of model learning. It is realized that the single pedestal model can generate pictures with rich details and texture on the premise of the integrity of the body. At the same time, the pedestal model also realizes the graphic ability of inputting pictures and outputting pictures with rich details.

Third, the transformation of various styles. The graphable large model has an automatic learning model based on Prompt, which provides users with different style templates based on the understanding and expansion of knowledge. According to the prompt word automatic expansion module, we can enrich the user description, including style, composition, visual elements and so on. Combined with the powerful text base model, Kolors can help users accurately understand their needs and generate pictures of various styles through a simple description.

Behind the powerful image generation ability of large models is the long-term accumulation of multi-modal content understanding and generation by Kuaishou AI team, as well as continuous innovative exploration in the field of large models.

It is understood that since late August, Kuaishou AI team has started internal testing of the Ketu large model platform within the company, providing internal business teams with rich and comprehensive AI painting creation capabilities, and supporting web-based tools and standardized API.

Kuaishou AI team said that the "graphable" large model continues to optimize results and enrich capabilities. In the future, Ketu will cooperate with more businesses of Kuaishou, enrich the imagination of the business with AI technology, explore more interesting and valuable innovative applications, and make man-machine co-creation a reality.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.