Is the AI technology that is advancing amidst the controversy really integrated into the development process of the game industry?

Source: Touch Music, Author: Zhu Siqi

**People are still trying to find the correct location. **

In recent years, AI has sparked a lot of topics and debates. In September last year, Touchle published the first article discussing AI painting. Since then, AI technology has come a long way. Not only drawing, AI can participate in text, dubbing and even soundtrack. The iteration of open source software and the emergence of many powerful plug-ins have also further improved the ease of use of AI tools.

At the same time, voices of controversy and anxiety around AI have never been far away. Not long ago, it was reported that Rhea Games fired the painter and completely replaced it with AI drawing. Rhea quickly issued a statement denying it, but players still questioned it due to the decline in the quality of the game card surface and other reasons. On the other hand, senior practitioners in the original painting industry released the results of a questionnaire survey, which summarized the requirements of multiple manufacturers for job seekers’ AI skills when recruiting, and also investigated the efficiency of painters at different levels after using AI. degree of improvement. This questionnaire is mainly to explain that art practitioners are still quite irreplaceable at this stage, and to alleviate some of the spreading panic.

Rayark games were once known for their unique art style, but now they are in controversy because of the suspicion of using AI instead of painters

This state of confusion is inevitable. When the emergence of AI is superimposed with the economic downturn and even the austerity of the entire industry, it is difficult for people to distinguish the proportion of AI factors in the wave of unemployment, and whether this technology is as omnipotent as rumored. Therefore, Touchle once again contacted the producers and visual directors of the game company, and even members of the AI team who switched jobs from the game industry to try to understand the current stage of the application of AI in the industry.

We eventually found that even though the AI function seems to be so powerful, there are still great differences in the application of it by different manufacturers and developers, and not all of them are integrated into the workflow-depending on the size of the manufacturer, the type of game product, and even Developers’ own understanding of AI technology and even game development itself.

Current Status of AI Applications

“How to integrate AI into the workflow?” A game developer’s answer was unexpected: “Good question, we also want to know, so we are preparing for a meeting to discuss.”

In a sense, this is the status quo of quite a few game manufacturers. The wave of AI has indeed come, but many people are still testing at the water’s edge.

Yoshikawa Mingjing is currently a member of an AI team, mainly providing AI drawing services for enterprises, as well as training in software technology. The team was established in early 2023 and is very optimistic about the prospects of the rapidly developing AI technology. At present, people in the team have different working backgrounds. Yoshikawa Mingjing is a game producer, so he has more contact with game manufacturers.

In the process, he found that the popularity of AI within different manufacturers is not the same. Big companies and independent developers like Tencent and Netease are very active in the application of AI technology. They are at opposite ends of the spectrum, wanting to put similar technology into use as quickly as possible for completely different reasons. A considerable number of manufacturers in the middle are still in the groping period, looking for ways to really improve efficiency. In other words, AI technology has not yet been fully integrated into the workflow.

The examples of big manufacturers can be obtained from their official case sharing published online. Since the spring of this year, teaching documents shared by top manufacturers have sporadically appeared on the Internet. It includes a fairly detailed StableDiffusion introductory manual, as well as some practical application cases.

In the summary of the AIGC workflow published by Netease, it can be seen that manufacturers have tried to apply AI technology to many fields such as CG business, event business, and virtual characters. This process is mainly carried out by humans for the initial requirements analysis and final polishing. In the middle, the search for references, the generation of design proposals through drafts, and even the partial refinement stages can involve AI.

In April of this year, Netease shared some of the AI-involved workflows, covering content not limited to game development

Judging from the cases in the summary, in terms of graphics, AI graphics are most often used in icons, live broadcast special effects, etc., which are usually not carefully examined by users, and the accuracy requirements are not particularly high, such as repeated treasure chest icons and medals in games. icons and more. The second is to quickly produce a variety of solutions in the early stage of scene design, fashion design and other applications, and select them for the demand side after preliminary refinement. Finally, there are some miscellaneous items, such as generating realistic virtual faces, simplified Q-version images, and even unifying the style of makeup photos of e-sports players.

In addition to mapping, ChatGPT is currently used to generate user questionnaire outlines and perform data analysis on existing data texts.

The historian of the game producer provided an application case of the division of labor for medium-sized manufacturers.

Due to the advancement of image generation technology and the emergence of the plug-in ControlNet that allows the binding of bones to adjust the screen, the process of drafting the original drawing for AI refinement is becoming more and more convenient. Selecting useful ones from a large number of AI refinement results for manual refinement can shorten the original 10-day workload to three or four days. “In fact, it liberates the manual work from digging out the details, and hand over the repetitive work and open work to AI.” The historian said.

In terms of programs, the programmers of Historian have basically developed the habit of “asking GPT for anything”, treating it as an advanced search engine, and asking GPT if they are unclear about any algorithm and method. The efficiency improvement of this move has achieved a leap forward with the evolution from GPT3.5 to GPT4. Even if the code given by GPT has bugs, you can throw it back and let it fix it yourself, which is equivalent to assigning a hard-working assistant to the programmer, and there is almost no additional cost.

Planning may also use GPT in the plot, such as setting an NPC’s character for the GPT thread, and then throwing the lines that the NPC wants to say to GPT, and let it speak with the specified character. This can avoid the situation where the copywriter cannot reflect enough personality differences when writing the dialogue.

Brother Peng, the visual director of a manufacturer, said that in addition to painting, AIGC also performed well in the field of video. AI voice overs and soundtracks are already fully ready for video creatives.

As for smaller independent teams and even individual developers, their application of AI may be more radical. On the one hand, these teams may be understaffed, and cost reduction is just needed. Whether AI can be used sometimes even determines whether they can enable a function. Some independent developers told Touchle that they could not afford to use original paintings, but with AI, some pictures can show the effect similar to that of a large company. The evolution of ChatGPT gives some individuals with no programming experience the opportunity to start trying to develop games.

In addition, independent developers may try experimental works according to their personal interests, regardless of trial and error. Now on Steam, there are sporadic pictures, scripts, dubbing and other games that are all generated by AI. They may not have a high level of play, and they cannot even be regarded as formal works, but they are also a direction for applying this type of technology.

** Obstacles still exist **

Many examples show that AI is available, but compared with some exaggerated remarks and effects on the Internet, it depends on many other variables to really improve the output of actual work. For example, the legendary “cost reduction and efficiency increase”, the actual situation is not the same as imagined.

According to Peng Ge’s feedback, in terms of “efficiency enhancement”, such as the drawing of characters and scene graphs, the average efficiency can be increased by more than 50% under the condition of AI, which is indeed a significant effect. However, in terms of “cost reduction”, the company’s cost savings are mainly saved on outsourcing, and internal costs are actually significantly higher, because more manpower is required to be responsible for the use and exploration of AI. “The things generated by AI now are still different from what you describe and imagine, especially the details cannot be as good as those directly designed by humans, so it will take more time for trial and error and repeated revisions.”

The historian also feels that there is still room for improvement in the accuracy of the feedback given by ChatGPT and others. “It is not yet possible to give particularly complete and systematic feedback. If you can hand in hand and give accurate code that can achieve a certain function, the difficulty of development for individual developers will be further reduced.”

“AI is really easy to get started, so it will give you an illusion that it is easy to use at first, and it is very good.” Brother Peng described his feelings about using AI in his work, "But if you want to apply it in depth and accurately in your work, you still have to There is a high threshold. It not only requires the professional ability of art, but also has good logical thinking ability.”

This mainly means that the operator needs to know what to let the AI generate, and the initial creativity and ideas are the most important. “To put it bluntly, AI can help you do a good job from 1 to 100, but it still depends on people from 0 to 1.”

Fully integrating AI into the workflow also needs to be explored. "For example, if an original painting is drawn by an artist from scratch, there will be multiple stages of draft, line draft, and color draft, and there will be many layers in between. Some companies have express regulations on this point, which is convenient The people who took over the manuscript before and after you got started smoothly. If you want to revise, you can clearly go back to a certain intermediate stage to revise.” Yoshikawa Mingjing said, “In addition, many characters have the need to do Live2D, and the draft needs to be cut. Points. However, the AI is directly thrown to you with a single layer, and it is actually very troublesome to modify it on this basis.”

Integrating AI tools into the workflow is equivalent to making it work well with humans

Because it involves modification, the degree to which AI improves the efficiency of painters of different levels is not the same. In the survey mentioned at the beginning of this article, downstream and upstream painters tend to believe that AI does not significantly improve efficiency. Because even if the downstream painters see the problem of AI drawing, the efficiency improvement is still not great because of the lack of modification ability. The upstream painters think that there are too many problems and the modification effect is too poor, so it is better to repaint. Most of the middle-level painters think that the efficiency has been improved significantly.

From the perspective of the team as a whole, in addition to the efficiency of output itself, improving communication efficiency is actually very important. "Because in the past, after brainstorming, it may take a few weeks for everyone to have a concept, and then implement this concept into a visual. Now it is very fast, and you may be able to come up with a very close to the finished product in a few hours. Let the demand side or the boss choose things.” Yoshikawa Mingjing said, “Many people will ignore the cost of such communication, but in fact, usually, such communication takes up a lot of everyone’s time.”

And Brother Peng believes that the essential reason for not knowing how to integrate into the process is that AI has revolutionized the production methods of game manufacturers. Just like when Photoshop and digital painting appeared in the first place, people should fundamentally consider changing the way of creation.

“For example, we went to outsourced drawing. There is a team in the outsourcing company. The chief artist is responsible for designing sketches, and the assistants are responsible for coloring. The painter who is good at painting faces specializes in painting faces, and then there are special people who dismantle parts and complete details, or Binding bones to generate Live2D and the like…is the division of labor of such a set of assembly lines.” Peng Ge said, “Now with AI, AI will give you a picture directly, and this process may be reversed. People are responsible for cutting out images, layering, and completing and modifying details…” Compared with some outsourcing agencies with varying standards, the quality of AI image output may be more stable.

Although both seem to be physical work, the latter currently takes less time than the former. “Originally, if you manually find the direction and refine it manually, even if you divide the layers at the beginning, it will take about 10 days in total. Now AI finds the direction for 1 day, AI refines it for 1 day, manually uses AI semi-finished products for 2 days, and manually refines it later. 2 days, a total of 6 days, which is still shortened compared to the original.” The historian calculated the overall efficiency, “And the two days when the AI is responsible for the work, the human is doing other things, or simply leaves work, and throws it to the graphics card. Count at night, without affecting each other.”

Of course, AI is not omnipotent. Several interviewees mentioned a typical example: AI still can’t help much with products that currently feature unique art styles.

One reason is caution. Manufacturers of a certain scale rarely use AI-generated content directly in commercial games, nor do they advertise this to players and users. “There are still some people who are very disgusted with this. The proportion may not be high, but the absolute number is not small.” Yoshikawa Mingjing said, “Especially in two-dimensional games, it is easy to be ‘flamed’ by players.” The value often lies in the art resources. Whether it is a real artist who “overturns” or really uses AI, he will face doubts.

Another reason is operability. Last year, some developers proposed that they want to train AI that adapts to their own style locally. According to Akiyoshi Yoshikawa, this problem, as well as the compatibility issues of AI tools and other office software that have been mentioned sporadically, have almost been solved technically. Today, not only commonly used tools such as Photoshop have launched built-in AI plug-ins, ChatGPT is connected to search engines, open source software such as StableDiffusion has allowed manufacturers to split and specialize functions according to their own needs. Theoretically speaking, each manufacturer can now have a set of AI tools tailored for its own project, and can train locally for its own product style.

Adobe Photoshop launched the official built-in AI function, which surprised design practitioners

But this tool set is not for everyone. Peng Ge explained the invisible resources needed by AI tools from the perspective of medium-sized manufacturers.

"For example, the future development direction of StableDiffision should mainly rely on B-end companies with redevelopment capabilities to make customized optimization adjustments and cultivate more unique models. But, this thing needs to rely on two foundations: one is computing power , and computing power means money; the other is data, and you have to have a lot of local data that can be fed to AI.”

Large companies must have more data than small and medium companies. "It’s like the advertising industry. Large advertising companies (such as 4A advertising) do hundreds of similar projects around the world every year, while small advertising companies (such as Creative Hot Shop) only have a few cases a year. Although one or two may do well It’s particularly good, and they’re all high-quality products, but the amount of data is difficult to support the directional training of AI.” Peng Ge said. Without directed training, AI mapping would be very homogeneous. “To put it bluntly, it is easy to train the most common and standard two-dimensional painting style that is circulated on the Internet, and there is no competitiveness.”

Therefore, in Peng Ge’s view, the rapid development of AI technology has indeed “decentralized” productivity, allowing small teams that cannot afford original paintings to have decent art, and allowing independent developers who can’t program to copy from ChatGPT. code, but the important thing is that big manufacturers still have leading advantages. “In order to develop AIGC in the future, the basic data accumulated in the previous era may be wealth.”

Looking for the correct location

Judging from the feedback from these developers, AI can already play a considerable role in the development process, but it is not as omnipotent and powerful as some people imagined or boasted.

The historian feels that if the effect of using AI in the team is not good, it may be that the manager has not thought about what the AI will do. This is ultimately a matter of perception.

"Some people hope that AI can achieve the output of finished products from scratch, which is obviously not possible at present. When they find that AI cannot directly give the most satisfactory results, and does not completely replace a certain manual function, they feel that AI is useless. "The historian said, “This idea of reaching the sky in one step is wrong. AI is an acceleration tool for developers. If you don’t recognize this, you can’t use AI well.” In other words, AI can’t even be counted as an invisible or intelligent employee. , but the assistant of the existing employees, which is used to improve the efficiency of the right people.

Brother Peng’s idea is that after the popularization of open source tools like StableDiffusion that allow manufacturers to customize, future AI software should be configured by large manufacturers or professional teams like office software like Feishu and DingTalk. Manufacturers to buy and use. Yoshikawa’s team has begun to consider providing similar services in long-term cooperation projects.

We are indeed on the brink of a new era, but we are far from actually stepping into it. Where the technology is used, whether it is radically used, and whether it is accepted by the audience, a large part of these are still waiting to be explored and verified.

At the end of the day, when “using AI” becomes the norm, people in the industry are trying to find the right position-both the position of AI as a development tool and the position of themselves as users. This will be a long period of exploration. It may still take time to tell whether opportunities or tougher challenges await developers in the end during this period.

(At the request of the interviewee, some characters in the article are pseudonyms.)