Useful, fun, futile or catastrophic… “Machine learning” excites the curiosity and creativity of manga publishers and animation producers. In the third part of our investigation, we looked at the various incursions of AI in support of the production teams of Japanese publishers.
Episode #3: AI at the service of production
If the creation of a manga is the essential step in the ecosystem of Japanese entertainment, it is not the only one to attract researchers in artificial intelligence. Several professions outside the creative field are already benefiting from technological advances – and others are considering it.
The companies Kodansha, Shogakukan and Shueisha joined in May 2021 with the company Marubeni to develop an AI to improve the distribution process. In recent years, the manga return rate has risen to 40%, which represents very high costs, and is undeniably an area for improvement. Even if the manga market has continued to climb in recent years (both in Japan and for export), publishers have an eye on indicators other than turnover or the number of manga sold. and works to improve their efficiency and ultimately their margins at each stage.
Another crucial point, on which the FR editors are also working. Reduce the number of returns/pillons and improve stock management. With digital prints we can be more reactive and above all reprint smaller quantities (from 30ex). pic.twitter.com/2XsFUjV25A
— Valentin Paquot (@Athan0r) May 30, 2022
In the new AI-assisted system being developed, partner bookstores will be equipped with RFID readers. Each manga will be assigned a chip, with a unique electronic identifier. This will make it possible to manage stocks in real time and improve restocking, without having to centralize requests. An overflow in an X library can even be used to re-store the Y library. It will eventually be possible to trigger reprint phases on demand based on so-called predictive algorithms. For example, by identifying places that are precursors to trends and anticipating a wave of manga purchases and reprinting it before it is sold out. Or react on a 2 or 3 day rolling selling trend instead of waiting for weekly or monthly numbers to ensure better responsiveness.
In 2018, the publisher Square Enix signed a partnership with the Chuo University research laboratory, which covers, among other things, the subjects of translation and artificial intelligence. Partnership extended in 2021 to the Matsuo laboratory of the University of Tokyo for research on the deep learning. Since then, the gross margin has exploded, impossible to know the impact of these partnerships, but they are probably not zero.
Moreover, the foreign market is no longer anecdotal for manga sales. However, the latter has recently been taken by storm by the webtoon. the korean wave, hallyu, surged and surprised Japanese publishers with its success. Faced with this phenomenon of full-color webtoons – and to conquer another type of reader – more and more manga are benefiting from a colorized version. The romantic comedy The Quintessential Quintuplet, big success on the Japanese archipelago, published in France by Pika edition, comes out in this new version colorized by a professional. But if profitability seems guaranteed for a hit, it is less certain for a more confidential title. Therefore, automated colorization solutions via Machine Learning are attracting the covetousness of publishers.
On a 100% automated model like the tool manga-colorization by Maksim Golyadkin and Ilya Makarov, “it’s a little flashy, it’s not super harmonious in terms of the colors and it’s going to be hard to get it to take into account elements that are out of box or not in the drawing. But the result, without human intervention, is impressive”, analyzes the professional colorist Yoann Guillo.
However, advanced support solutions are very good, such as community software Krytaor works of hong kong university research allow colorists to save a huge amount of time by manually applying color to an entire “area” of the image with a single brushstroke. Indeed, the AI has learned to interpret the sections of a drawing, and even to distinguish the intensity of the line to apply more or less rich tints. The latter can even take into account the number of lines or the types of frames. “Solutions like Kryta, which is less complete than Photoshop, allow you to rough out the work before attacking the goldsmithing phase in Photoshop”, concludes Yoann Guillo.
AI is also useful for automatically selecting content from the rich catalog available to major Japanese publishing houses. La Shueisha has just published three artbooks grouping exclusively portraits of the characters of One Piece. The particularity of these books lies in the fact that an AI was trained by “machine learning” to recognize and extract the faces of the characters created by Eiichiro Oda. The first 60 volumes of the famous manga were analyzed to isolate 74,121 faces and their expressions. A titanic work that would have made more than one graphic designer cry!
This first compilation allows us to fully appreciate not only the evolution of Oda’s line but also the richness of the emotions he illustrates through his characters. The publisher took the opportunity to reveal that the author had drawn by his own hand more than 100,000 original characters during the first 25 years of publication of the saga. One Piece. A figure to make you dizzy…
Beyond this example of three artbooks, partly generated using AI, there are several other types of books that could benefit from such algorithms. Guide books or school diaries are two obvious examples, as their creation requires a substantial work of research and qualification of the manga content. Many of these processes could benefit from being automated in part through well-trained machine learning.
Another area where AI is surprisingly useful is… censorship. The majority of publishers have launched their manga reading application on smartphones or tablets. Only downside: these applications are generally restricted to the Japanese or even Korean market. But the international market now representing for some publishers nearly 30% of their income, it is impossible to cut oneself off from this potential windfall. More and more of them are putting new applications online to make their catalog accessible to foreign readers. Most often in English, some sometimes support French (like Manga Plus from the publisher Shueisha). Publishers have chosen to apply censorship via “machine learning” on manga boards which could cause the rejection of the application in certain countries where the legislation is very strict in order to keep their applications worldwide. For the moment, these algorithms for recognizing areas to be censored are a little too vehement both in content and in form (large black band). The publishers will surely review the algorithms or even remove the application in certain countries.
Predictive algorithms should also be deployed to track and analyze market trends. For example, by monitoring the themes of the sales of Light Novels, to refine the analyzes on the titles to be adapted into manga or anime as soon as possible. Ditto for competitive intelligence.
Finally, for the curation of available content, since 2017 the publisher Kodansha has approached the pure player Digital Garage, specializing in user-generated content. Together, they have developed a “machine learning” system that observes and learns user trends via social networks to support the editorial teams of women’s magazines to offer personalized content to readers who love fashion, beauty, lifestyle, interior design – or more broadly entertainment.
Another tight flow Japanese cultural market: animation. Perhaps even more desperate than that of the manga, as the working conditions are so hard for a salary gratification rarely at the level of the efforts. In the next part of our investigation, we will see where and under what conditions “machine learning” is being studied in this merciless universe…