At MIXI, Inc., we are working to create new entertainment experiences by combining AI technology with character data and other game data that we have accumulated through many years of operating MONSTER STRIKE.
The AI Modeling Group of the Development Department’s Tanpopo Division leverages AI and machine learning expertise to test and utilize the technology while collaborating with other business units. As part of these efforts, MONSTER STRIKE has introduced the in-game content “Quiz DE Strike” and the experiential content “AI Monster Ball Maker” for the offline event DREAMDAZE IV, both of which make use of generative AI.

Quiz DE Strike automatically generates hints from massive amounts of character data

Quiz DE Strike is in-game content where players guess MONSTER STRIKE characters based on multiple hints. Players guess the correct character with as few hints as possible while receiving hints of varying difficulty levels one by one.

Hint generation integrating varied character data

Since there are a vast number of characters in MONSTER STRIKE, creating varied hints for each one would require a significant amount of time. Therefore, we built a system that automatically generates hints using an LLM¹ based on official character backstories, in-game specifications, character illustrations, and more. We generate hints in the categories of specifications, backstory, and appearance, and set a difficulty level for each. In addition to character specifications and backstories, we also use AI to translate visual information from character illustrations into text, allowing us to use visual characteristics from images for hint generation as well.

Generating hints while comparing against other characters

If only information about a single character is provided to the LLM, it is easy for hints that apply broadly to other characters to be generated. By providing the LLM with information about other characters in addition to the target character, we are able to take differences between them into account and generate hints that are more unique to the specific character in question.
Furthermore, we estimate how difficult it is to identify the character based on each hint, and, by combining questions of varying difficulty levels, we have designed the game so that hints can be revealed in stages.

The AI Monster Ball Maker allows guests to become MONSTER STRIKE characters

At the MONSTER STRIKE offline event DREAMDAZE Ⅳ held on July 11 and 12, 2026, we provided the AI Monster Ball Maker, a system that uses AI to generate illustrations in the style of MONSTER STRIKE’s monster balls based on photos of guests. The generated images were used in animations mimicking the in-game Hatcher, as well as mini-games at the venue. They were also provided to guests as acrylic keychains.

Generative AI technology that reproduces MONSTER STRIKE’s art style

To generate monster ball illustrations, which are drawn in a style that is unique to MONSTER STRIKE, we built a generation pipeline that reproduces this style by utilizing all of the character images and monster ball illustrations produced to date for additional training of the image generation model. In this way, we are leveraging technology honed through our internal technical testing of image generation AI to deliver real-time experiences for guests.
Instead of converting photos directly into monster ball illustrations, we first use a general-purpose image generation AI to convert them into anime-style illustrations while maintaining features such as the person’s face, hairstyle, clothing, and pose. Then, we use a model trained on MONSTER STRIKE monster ball illustrations to convert them into that style. Finally, we use another image generation AI to correct localized flaws such as fingers and line art. By combining the strengths of general-purpose models and specialized models, we aimed to achieve both a convincing likeness and the distinctive style used in MONSTER STRIKE’s monster ball illustrations.

Quality verification and automatic regeneration by AI

With generative AI, issues such as distorted fingers and other body parts, facial features that are different from the original person’s, and differences in pose may occur. Since it is difficult to manually check all generated images at an offline event, we incorporated a quality check using a VLM (Vision-Language Model) into the generation process.
We compare the original photo with the generated image to ensure that the person’s features, expression, pose, and the overall composition have not changed significantly, and we also detect flaws such as distorted fingers or other body parts in the completed monster ball illustration. If a problem is detected, the process returns to the relevant step and automatically regenerates the image.
In addition, for the depiction of fingers, which generative AI struggles with, we improved the accuracy of the model itself by adjusting prompts and adding training data based on real issues that occurred.

Built for stable operation during offline events

Since it is necessary to provide the experience to many guests in a short time at an offline event, we built a system that can handle not only high image generation accuracy but also intermittent downtime of external AI or GPU² resources. The system is equipped to execute image generation processing asynchronously, manage error and retry statuses, provide fallbacks when GPUs cannot be secured, and monitor processing times and error counts. During the two-day event, 955 images were generated, with 99.8% completing successfully. Processing time was also between 71 and 73 seconds in about 90% of cases, achieving a level where each guest’s image generation is completed while the previous guest is playing the mini-game.

Combining AI with proprietary data and technology to create new experiences

For Quiz DE Strike, we used AI to enable large-scale content production that makes use of information about a vast number of characters that has built up over many years. For the AI Monster Ball Maker, we created an experience that allows visitors to immerse themselves in the world of MONSTER by using a model trained on MONSTER STRIKE’s art style together with generative AI.
At MIXI, we remain committed to making use of the proprietary data and creative assets of each of our services in combination with AI technology. This is not only to enhance our development and production processes but also to create innovative entertainment experiences and deliver new value.