Wall Street Insight Research specially invited Founder Securities AI Internet chief analyst [Yang Xiaofeng] to disassemble in detail what impact AI large models can bring to the game industry to reduce costs and increase efficiency, and preliminarily predict that the impact will take at least two quarters to be reflected financially.
Core Ideas:
1. AI now has two large models that can greatly improve the efficiency of the game in the art production process, save production costs and shorten time. The two large models, Stable Diffusion and NeRF, are optimized for 2D and 3D scenes and character generation during game art production, respectively.
2, Stable Diffusion, also known as the Wensheng graph model, can generate multi-perspective 2D maps, further laying the foundation for the generation of 3D maps, but the conditions are limited by the need for high-performance graphics cards, so it is difficult to penetrate widely in the general consumer group.
3. The NeRF model mainly reduces the cost and efficiency of the 2D-3D process, and the current cycle of making a 3D game character is about 30-45 days, and it requires more steps and many participants; Using NeRF allows for rapid modeling, resulting in increased efficiency and cost savings. Work that used to take 10 working days can now be completed in half a day to a day, which equates to a 90% time savings.
4, but the NeRF model has not yet reached the tipping point, the main reason is that the technology has certain barriers, but the leading company Luma AI, developed NeRF-related APP, has been launched APP Store, greatly reducing the threshold for the use of NeRF, is expected to make breakthroughs in the next six months to a year.
5. In general, AI models can greatly optimize the cost and time in the game development and design process, according to incomplete statistics, the art cost generally accounts for 50% to 80% of the game development cost. If a game company's R&D costs account for 70%, then 40% of them can basically be greatly reduced.
6, the difference between large companies and small companies is whether they can only reduce 2D money, or can also reduce 3D money, if 2D and 3D can be reduced, then the entire cost reduction and efficiency increase is actually 60%-70% is not necessarily able to fight, so the decline is actually very large.
Body:
See Intelligence Research: Why is the application of AI in games getting attention?
Yang Xiaofeng:
The core reason is that the business model of game companies is relatively excellent, and they have certain resources to use AI models. Another reason is that there are now many AI models that are relatively mature, such as AI painting, which directly subverts or changes the game industry, so the game industry is currently the fastest industry to apply AI. Of course, the animation industry is similar, because the level of technology at home and abroad is basically the same, especially in AI painting.
Recently, we noticed a game overseas in which several characters are controlled by artificial intelligence. As you can imagine, this game is like Westworld, with many characters in the game having their own growth trajectories and personalities. In this world, everyone is real except the players themselves. The immersion of this experience is very high, and this application will become more and more popular. This experience is actually an improvement on NPCs.
We also noticed another case where when we made art with games, for the same money, we could make the whole art scene more and more beautiful. In the past, we could only make the front perspective more beautiful, but now we can make the whole vision very beautiful for the same money. We've seen a lot of such cases.
In addition, we recently discovered a platform called Inword. This product has been trained in advance to various very personal characters, and this code can be directly integrated into the game. In this way, characters with character like Musk may appear in the game. Others have already integrated these, just need to access it, and the game experience will be improved very quickly.
See Intelligence Research: What AI models can be applied to the gaming field? What are the characteristics of each?
Yang Xiaofeng:
There are now two mature models, one is a large text model, the other is a 2D graphics model, and there is a 2D to 3D NeRF model that is in the explosion period.
The first type of model is the text generation model, such as ChatGPT and various large models in China, which can accept text input and output corresponding text, which is widely used. But in games, the characters in it are usually used directly.
The second type of model is similar to the Wensheng diagram, which can generate a variety of pictures based on the input text. This model is so mature that most internet companies are using it to generate pictures. The more famous are Stable Diffusion and Midjourney, of which Stable Diffusion is more widely used because it requires less cost; Midjourney is consumer-oriented, has a better user experience, and has lower requirements for computers. Therefore, it can be used when making 2D images later.
Now there's also a model that is about to explode, likely in the next six months to a year, and it's a model that converts 2D to 3D called the Neural Radiation Field (NeRF). Just take a few photos and instantly generate a 3D model that includes scenes and people. Imagine scenes that were previously needed to make games, such as the Paris scene from Assassin's Creed. Now it only takes some money to shoot some video with a drone, and the 3D model can be built in no time, and the cost is reduced very quickly. This is the model that may already be in use today.
See Intelligence Research: What are the ways to generate 3D models? What are the advantages and disadvantages of each?
Yang Xiaofeng:
In the past, the method of generating 3D was generally modeled by drawing floor plans, for example, for a character, it may be necessary to draw multiple drawings from different perspectives to model the model. The modeler will then use modeling software to build one by one, such as sticking three-dimensional hair to a human face. This method takes a lot of time, and can take 2 to 3 weeks to model a character. The advantage is that each object is a relatively elaborate three-dimensional structure, but the disadvantage is that the time cost is high, so it is generally handed over to the outsourcing company to complete.
There is also a method called photographic technology, that is, to build a model by taking photos of objects, but each photo must coincide by 50% to build a model, but this technique is difficult to process details such as light and shadow, so it is rarely used.
The recently used NeRF model has been published in a paper, and only two software are currently available: Instant NGP and Luma. Both products are now capable of instantly generating 3D models by photographing just one object. But it also has the disadvantage that compatibility is not considered when making products.
For example, we made a very beautiful 3D model within the NVIDIA ecosystem, but we wouldn't consider optimizing it and running it in Unity or Unreal Engine, because once the 3D modeling was done, it had to be put into the game engine to produce the best results. It seems that this area is still evolving, but Luma has put its plugins into Unreal Engine and game engines, and it could progress even faster. We can understand it this way: the current model is not very open source, so it will take some AI engineers to use it better. The advantage is that it can indeed reduce a lot of costs, but it requires some relatively high level of skills.
See Intelligence Research: Will NeRF replace the original 3D model as the mainstream choice in the future?
Yang Xiaofeng:
I think there's a good chance it will happen, because the latest developments in the industry right now are, as we just described, 2D photos can generate 3D models. Now, the latest developments in the industry can be directly modified, for example, I have a 3D model of my own character, and I can replace my avatar with Musk's avatar by entering text. This technology is quite mature.
Recently, there was an overseas freshman who contributed very quickly to NeRF technology. In the future, you can modify the 3D model with text, for example, you can have it add a beard or glasses to your model. This feature looks cool and quite powerful, but if you understand how it works, you'll see that it's not that hard, it's just that no one has optimized it before.
Most people nowadays use models for grafting, each with its own area of expertise. For example, text models excel at processing text input and output, Stable Diffusion excels at generating 2D images, and NeRF models excel at converting 2D images to 3D images.
The future trend is to connect all models together, so that users can simply express their intent, and the model can help achieve the task. Currently, the NeRF model is one of the most promising, but it has not yet reached its tipping point. The NeRF model is expected to shine in the next six months to a year.
See Intelligence Research: What are the characteristics of the Stable Diffusion model?
Yang Xiaofeng:
We just mentioned that NERF can convert 2D photos into 3D images, while Stable Diffusion is essentially a tool for literary diagrams. Why is this tool so popular? How powerful is it at the moment?
Before November and December last year, the software was not really hot, but it suddenly became popular. The reason is that someone uploads a data package, and the user only needs to enter the desired cartoon or real person, and a very beautiful image can be quickly generated, attracting a large influx of users. At the same time, countless people have also begun to provide various training data packages for it, further enriching its functions, and now can draw not only two-dimensional cartoons and real people, but also GTA5 images.
What makes this software so powerful is that countless users in the market are providing it with training data packages that enable models to generate all kinds of beautiful images. However, it should be noted that this software needs a better graphics card, preferably 3090 or higher 4090, so the user experience for ordinary consumers may be average. To this end, someone has specially developed a web version using Stable Diffusion's modeling, so that users do not need to have a good computer to use it on the web. However, there is a fee for this service, which is tens of dollars per month for regular users and higher for business users.
Because of open source, Stable Diffusion is now capable of generating not only still images, but even video. This is because in March of this year, someone modified the underlying code to make it possible to specify some parameters for AI mapping, for example, if I draw a horse, I can draw another image of the horse raising its legs, and then raise each horse's legs into one video. This is why many companies at home and abroad have suddenly launched AI videos, because they are all based on the principle of Stable Diffusion, but there may be some optimization on this channel.
In addition, Stable Diffusion is now more powerful in the ability to build 3D models out of nothing, as long as you can describe the angle of the object, it can generate 2D images from multiple angles, and then use the multi-view 2D images to generate 3D images. Therefore, in the future, 3D can be generated out of nothing. This means that Stable Diffusion has become one of the most powerful tools in the field of AI mapping, and other tools are iterating on top of it.
See Wisdom Research: What are the advantages and significance of the open source model? What does the extension plugin do for the model?
Yang Xiaofeng:
Once the model is open source, the world's top talents can use it and fine-tune it, and they can contribute a variety of asset packs to make the model more powerful. Because Stable Diffusion is open source, everyone in the world can enjoy the dividends of this AI graphing, and everyone can modify it on it or put it on their own servers. Open source means that every piece of code in this software is publicly available and can be downloaded locally, and no one else can manipulate you.
Of course, the industry may need some relatively powerful talents to push this product to a higher level, modify the underlying code, and further improve the product level. Therefore, open source is progressing to the whole model very quickly, and it is conceivable that before November and December, most people were at a relatively low level, but because of open source, countless people uploaded data packages in January and February, and modified the underlying code in March, and the progress of this product is very fast. It can be understood that open source has raised the level of everyone to a very high level.
See Intelligence Research: Model open source greatly accelerates the implementation of the application level?
Yang Xiaofeng:
Why is Midjourney so popular? Not because it has a technical advantage, but because it is more able to meet the needs of C-end users. For example, we all know that Stable Diffusion technology is good and the product performance is good, but the problem is that not everyone's computer has such a high graphics card, and at the same time, more data packets does not necessarily mean a better experience, because many people want something more realistic, atmospheric effect, right? Therefore, many products are optimized for the experience of C-end users on this basis. The technology behind it may require the use of original technology such as Stable Diffusion, or other technologies for cost considerations, but for the average C-end user, using Midjourney is basically enough.
See Intelligence Research: How does AI reduce the cost and increase the efficiency of games?
Yang Xiaofeng:
For example, a game company typically accounts for 50% to 80% of the company's R&D costs. Because some games have a very large number of users and can not have problems such as lag, the program cost of the game is very high, then the art cost accounts for 50% of the research and development cost. But some games are just card games, just need to draw the cards, it can move on its own, this kind of game art costs account for 80%. Imagine the cost in this case.
As a game artist, first of all, I need to design the UI interface of the game, if I have 10 people on hand, one of them will be responsible for drawing the UI interface of the game's horizontal screen, such as the login screen. The remaining three people will design character models and do 2D character design, such as drawing characters from multiple perspectives, which usually takes two weeks to complete the drawing of a character.
What used to take these four people to complete in two weeks can now be completed in half a day. This means that we can save a lot of time and costs. Work that previously took 10 working days can now be completed in half a day to a day, which equates to a 90% time savings.
If we only consider the 2D aspect, it is the 2D character and flat interface. The remaining six layers are outsourced to someone else for modeling. For example, if I get a 2D image, I can get a modeler to help me build it into a 3D object or person, and the money is usually given to outsiders. This part of the cost can be cut.
But NeRF is not open source, and many companies don't have such good AI talent to use it, so progress in this area is not very big. If a game company's R&D costs account for 70%, then 40% of them can basically be greatly reduced.
So I think the difference between large companies and small companies is whether you can only reduce 2D money, or you can also reduce 3D money, if you 2D plus 3D can be reduced, then I think the entire cost reduction and efficiency increase is actually 60%-70% is not necessarily able to fight, so the decline is actually very large.
Interactive Session:
See Intelligence Research: How long will the results of game cost reduction and efficiency increase be reflected in the financial statements?
Yang Xiaofeng:
This mainly looks at a top-down logic, because the tool of 2D painting only began to reflect in January and February this year, and its implementation within the company will basically be after April, and may gradually appear in the second or even third quarters.
From the perspective of the model, the function of generating multiple perspective maps only began in March, and the second quarter slowly began to become proficient, and the third quarter may slowly show the finished effect.
Insight Research: What will be the impact on companies with more IP?
Yang Xiaofeng:
The value of IP is considered high because it can consistently produce products. By increasing production capacity, like Disney animation, supply can increase substantially, but demand may not keep up. AI can increase production capacity by up to 5 times, but whether the market can withstand so much is uncertain and may need to be discounted. Having a strong IP can increase the appeal of a product, as it becomes increasingly difficult to create new IP as the number of products increases.
See Wisdom Research: What are the focuses on the ways in which different types of game companies reduce costs and increase efficiency?
Yang Xiaofeng:
The first thing we look at is that for the head company, it reduces costs and increases efficiency particularly obviously, that is, 2D and 3D are reduced at the same time, because they recruit some top AI engineers, so 2D and 3D can be reduced. They may be able to compete with overseas 3A factories, because the kind of art barriers built by 3A factories in the past are relatively leaky, so this is the head company, then for the middle waist company, relatively speaking, it can reduce its own costs by using certain technology.
Wisdom Research: Which tracks are you optimistic about next?
Yang Xiaofeng:
AI painting is one of the more mature technologies at present, of which animation and games are the most likely to be the fastest field of application. Due to the previous shortage of art talent, these industries will greatly increase their production capacity after using AI to paint. In addition, the cost of painting with AI will also decrease significantly over time. Therefore, these two areas are worthy of attention, not only to reduce costs, but also to quickly increase production capacity.
![The AI large model is on the car, and Xpeng Zhijia continues to lead[fig]](data:image/gif;base64,R0lGODlhAQABAIAAAP///wAAACwAAAAAAQABAAACAkQBADs=)