What is the mechanism of 'Red Ram' that can generate mystery games by leaving it to AI? [CEDEC 2024]

A session held on August 23rd, the third day of CEDEC 2024``Almost a full AI game. A mechanism that allows Red Ram to generate an unlimited number of mystery games of the player's preference''.
"Red Ram" is a technology demo developed by Morikatron, a developer specializing in game AI, and is an attempt to automatically generate a murder mystery-style culprit guessing game based on keywords entered by the player. It has become.In this session, Shigenori Miyamoto, Morikatron's lead engineer who was in charge of natural language processing for the game, and Rikito Takahashi, an AI engineer who was in charge of system design and client implementation, explained how the generation works. It was revealed.
Lead engineer Shigenori Miyamoto (right) and AI engineer Rikito Takahashi (left)

Fun created by using generation AI in the game
The session started with Mr. Takahashi talking about the message he had put into this session. That is, ``I want more games to use generation AI in-game.'' This is a message unique to Morikatron, which is skilled in AI technology, but the company believes that by utilizing generative AI in games, two types of fun can be created.
One is that there are so many variations that you won't get bored. Another benefit is the ability to create a gaming experience tailored to each player's preferences. ``Red Ram'', which will be discussed in this session, was created to study the gaming experience unique to generative AI.

“Red Ram” game overview

So, how does Red Ram game generation work?First, the player inputs their preferred settings, and then the AI generates a plot that is the outline of the incident. Next, story data is created that fleshes out the data, and characters and settings are determined based on that data. Finally, dialogue lines, character images, background images, etc. are generated, and the game is completed by combining them.
Mr. Miyamoto explained the details of these generation. First, plots are generated in stages using LLM(large-scale language model/``Red Ram'' uses OpenAI's GPT 3.5 and 4) based on the keywords entered by the player. If complex conditions are required for generation, the high-performance GPT 4 is apparently used.
For example, even if the player inputs something that is not suitable as a weapon, LLM will create a scenario accordingly.


In plot generation, not only the incident setting but also a mystery writing guide and how to create a progress flow are entered as prompts. Prepare multiple progress flows that are manually generated by LLM in advance and use them randomly. It is said that generating plots without any format constraints can be expected to yield higher quality products.


From the plots created in this way, story data is generated in a format that can be handled by programs. As you can see in the slide below, at the plot stage it is just plain text, but necessary data such as locations and victims are extracted from it, and the content is fleshed out at the same time.


Next, detailed data for each person, location, and piece of evidence appearing in the game is generated from the story data. In addition to determining a person's profile such as age, personality, and occupation, their characteristic tone of voice is also determined. At this time, a prompt for generating the image is also created.


The last thing that is generated is a conversation scenario -- the text of the interrogation (conversation line) that the player, the detective, conducts on the suspect. By inserting the data generated so far into six types of templates depending on the situation and location of the incident, prompts are created and conversations are generated. In addition, the character's emotions are inferred from the text of the dialogue to LLM and classified into six types of emotions. This is said to be used to change the suspect's facial expression when generating images.


Next, move on to image generation. Based on the text generated by LLM, images are prepared using "Stable Diffusion"(the model name of the image generation AI). For the background and evidence images, use the prompts created when you generated the respective details. At this time, adjustments are made to adjust the viewing angle of the background image and to replace and draw evidence items that are difficult to represent.


Similarly, the character portraits use the prompts created when detailed data is generated. Constraints are applied so that the image is drawn in a fixed position and at a fixed size, and background transparency processing (cutting out) is also performed at this point.
For cropping, Semantic Segmentation (a method of segmentation that labels each pixel) and technology that estimates the depth of the image are used to ensure reliable cropping by extracting only the deep parts.

The facial expressions are adjusted so that they don't look like a different person by adding only a minimal amount to the prompts when generating the pose.

Challenges and future prospects for generative AI games

The next slide shows the incident that was actually generated as a result.


First of all, since the murder weapon was ``tofu'', it would be fair to say that the cause of death was suffocation. The stage was set to be an ``innovative game development company,'' so a place like an office or conference room was generated. Since the motive for the crime was"financial troubles"
, a human relationship centered around lending and borrowing money was created. By the way, it seems that you can choose between live-action style and anime style for the characters.






At the end of the session, Mr. Takahashi spoke about current issues and future prospects.First, there is the issue of quality. In "Red Ram," the characters' conversations sometimes don't make sense. For example, a conversation may begin with ``evidence has been found,'' but end in a mysterious manner, with the evidence not being found. What's more, the main character, the detective, may even reveal the truth during the explanation of the case at the beginning.This makes it impossible to enjoy solving murder cases and reasoning. In other words, the current situation is a game where you can enjoy the strange behavior of the generative AI. That has a certain flavor and is interesting, but that's not what Morikatron is aiming for.
In order to achieve a consistent scenario, it is necessary to further devise the system design and the prompts given to the LLM. Other issues such as cost and waiting time were also raised, but these are expected to be gradually resolved through the evolution of the LLM model.
According to Mr. Takahashi, in the future he would like to create an experience where the information entered by the player is reflected in the game world in real time, and in July 2024, he will release a technology demo called "Labyrinth of Kotodama". are.


This demo is a text adventure that aims to conquer a dungeon, but the system is such that items are generated based on the words entered by the player. In response to the input words, the item's image, parameters, flavor text, etc. are generated on the spot, and the main character is shown using the item.


Additionally, papers and generation examples related to "Red Ram" are available online. It can be viewed using the QR code on the slide below, so if you are interested, why not check it out?

The above is the detailed content of What is the mechanism of 'Red Ram' that can generate mystery games by leaving it to AI? [CEDEC 2024]. For more information, please follow other related articles on the PHP Chinese website!