Kuaishou internal beta AI play review: What is the collision effect between large models and short videos?

The "Battle of One Hundred Models" has recently added another participant. Following the launch of the large language model "Ruyi" of Wenshengwen last month, Kuaishou recently launched a self-developed large model in the field of "Wenshengtu". Figureable” (Kolors). As a short video platform, Kuaishou’s “Ketu” will naturally be used in its own App. Relying on the Ketu large model, Kuaishou has also begun to test the “AI play review” function in the short video comment area, trying to unlock the AIGC short video New ways to play.

It is reported that Kuaishou’s “AI Play Review” is the first time in the industry to apply AIGC capabilities in the comment area of ​​the core business scenario of a large-scale app. This function is designed to enhance users’ interactive experience in the comment area. Users can input creative text to Easily generate a large number of images in different styles to enrich comment interaction. Users only need to enter a text comment of 6 words or more in the comment area of ​​the short video, and click the "AI" logo in the lower right corner of the comment box to generate a comment picture with one click. They can also click "Change View" to switch to more styles. .

According to the Kuaishou AI team, through the "AI Play Review" function, users can express their opinions and emotions more accurately and more interestingly, and have more convenient and interesting interactions in the comment area, without having to look for suitable pictures. Or emoticon package, but can directly generate a picture. It is understood that AI game reviews can generate pictures ranging from common styles such as cyberpunk, pixels, and realistic animation, to pictures with strong personal styles such as Makoto Shinkai, Hayao Miyazaki, and Katsuhiro Otomo.

By analyzing the content input by the user, drawing semantic pictures has become a standard function of Stable Diffusio, midjourney, and various large AI models with Vincentian diagram functions in the domestic market. In other words, Kuaishou's AI review is essentially an AI painting tool. The technology behind it is mainly based on NLP natural semantic processing, and accurately identifying what the user wants to express is a key element

The effect of AI game review depends on the prompt word (Prompt). According to netizens’ experience, if the text comments contain more descriptive content about people, scenery, space, actions, etc., the generated pictures will be more consistent with the actual situation. On the contrary, if there are vague descriptions in the comments that lack specific referents, such as "666" or "Oh my god! Sister is so awesome!", the results generated by the AI ​​will not be viewable. Therefore, this reality directly leads to the fact that AI game reviews may not be loved by most users

The question is, what is the comment area of ​​the short video platform like at this stage? In fact, this is a scene full of witticisms, jokes, witty remarks and other emotional content. Due to the characteristics of short videos, including magical brainwashing background music, intensely stimulating pictures and uncertain reward mechanisms, users give up thinking and become immersed in it. Therefore, comments in the comment area are usually just a simple sentence, which users will use to clearly express their likes, dislikes or opinions

The result of this reality is that the content output by users in the short video comment area is basically emotional and lacks qualitative content. Just imagine, if it is just a pile of adjectives, AI will face the confusion of lacking a subject, which means that the final content generated by AI may be very different from what the user wants to express. I believe that friends who have used tools such as Stable Diffusio and midjourney know that if Prompt is mainly adjectives, the result of the lack of nouns is that the AI ​​will let itself go.

Even the most advanced GPT-4 is actually flawed in experiencing human emotions. In fact, AI's emotional perception ability is still a problem facing all AI researchers at this stage. At present, many large AI models are oriented to either serious productivity scenarios or conversations with humans, and almost no AI involves emotional expression. So in this way, it is actually difficult for Kuaishou’s AI game reviewers to do their job well. It might be good not to hinder users’ comments.

So in this case, why does Kuaishou launch AI game review? Of course, the purpose is to make the large model of Vincent's picture "pictureable" and have a realistic scene. The Kuaishou App itself is almost Kuaishou’s only consumer-oriented product, so “AIGC short video” has become almost the only card they can play. In fact, we can see from here that Kuaishou, as a new giant emerging in the mobile Internet era, is still inferior to traditional giants such as BAT in terms of background.

Unlike BAT, which has almost built itself into an Internet water, coal and electricity company, Kuaishou, a group of new giants that grew up in the mobile Internet era, currently almost all show the characteristics of a single business line of "strong trunks and weak branches", such as Kuaishou’s core business is basically based on the Kuaishou App, and almost all other businesses have not yet been launched. Before this round of AI concepts broke out, Baidu, which was once considered lonely by the outside world, in addition to a search engine, also made an input method, so Baidu's native AI applications can be carried on Baidu input method.

Looking back at Kuaishou, apart from the Kuaishou App, where else can the "tutu" large model be used? If Kuaishou wants to make an app solely for large AI models, Kuaishou may lose the opportunity. The current situation is that there is actually no generational difference in performance between the major AI models in the domestic market. The actual use experience of each model is basically the same, and the user's choice is often as long as it is useful. Even for users who want to experience the charm of large AI models, many have downloaded Baidu Wenxinyiyan, which has a first-mover advantage.

In fact, station B may have set a better example for combining AIGC with video. Previously this summer, Station B launched the "AI Video Assistant" account. Users only need to @AI Video Assistant in the comment area of ​​the corresponding video, and the latter can automatically generate a text summary of the video. For the long videos of Station B, the summary and organization of the AI ​​video assistant can help users complete information extraction in a short time, so it will naturally be welcomed by many users.

As a product with more prominent entertainment attributes, if Kuaishou App wants to better integrate with AIGC, it must naturally meet users’ entertainment needs. For example, intelligently generating emoticons based on comments may be far more suitable for the atmosphere of the platform than creating pictures of people in the comment area.

The above is the detailed content of Kuaishou internal beta AI play review: What is the collision effect between large models and short videos?. For more information, please follow other related articles on the PHP Chinese website!