Tokyo, May 20, 2025 — As of July 2024, Amana Images Inc. (Subsidiary of the Visual Bank Group) and AIST Solutions Inc. (Subsidiary of Japan’s National Institute of Advanced Industrial Science and Technology, AIST) entered into a joint-research agreement to develop a “Japan-made Image generation AI model with minimized rights risk”.
(See the press release here (available in Japanese only)
The “Japan-made image generation AI model with minimized rights risk” is a unique model that combines AIST’s formula-driven supervised learning, which uses figures generated from mathematical formulas instead of real images as training data, with real image data from the Qlean Dataset, managed by Amana Images, a Visual Bank Group company. The Qlean Dataset consists of images with clearly defined rights and no ethical issues. This approach eliminates concerns over rights infringement, which have recently become a major issue in the AI field, and enables safe commercial use.
This research aims to build an image generation AI model based on formula-driven supervised learning, extending to the development of vision language models as well. The outcomes are expected to promote broader AI utilization across industries.
In this article, we invited Seiji Osaka, Representative Director of AIST Solutions Inc., and Yoshio Tanaka, Director-General of the Information and Human Factors Domain at AIST, to join a discussion with Ippei Mochizuki, Representative Director of Amana Images Inc. (Visual Bank Group), and Shoma Akamatsu, Tech Lead at Visual Bank, to talk about the background of this joint research and the significance of developing the “Japan-made image generation AI model with minimized rights risk.”
Seiji Osaka
Representative Director of AIST Solutions
Joined TDK Corporation in 1982 and contributed to expanding overseas sales channels and global market share for the recording media business in Germany, Russia, and Italy. From 1999, engaged in corporate planning, overseeing divestment of the recording media business, acquisition of ATL’s lithium-ion battery business in China, Alps Electric’s HDD head business, Lambda’s power supply business, EPCOS’s electronic components business in Germany, and establishment of the RF360 joint venture with Qualcomm in the United States. Successfully led the restructuring of TDK’s business portfolio. Served as Senior Executive Officer and Head of Corporate Strategy from 2017. Departed TDK in March 2023 and assumed current role in April 2023.
Yoshiro Tanaka
Director-General, Information and Human Factors Domain, AIST
Earned degree from Keio University, followed by experience at a research consortium before joining the Electrotechnical Laboratory in 2000. From 2001, engaged in grid computing research at AIST, collaborating with overseas institutions to develop standard security policies for shared use of supercomputers, experimental equipment, and research data worldwide. Gained expertise in security through supercomputer operations and related projects. Served as Director of the Security and Information Promotion Department before appointment to current role in 2023.
Ippei Mochizuki
Representative Director of Amana images inc. (Visual Bank Group)
Graduated from the University of Tokyo Faculty of Law and joined SMBC Nikko Securities, working in investment banking for six years across M&A and fundraising projects in both traditional and advanced technology sectors. Joined Amana Images in August 2022, responsible for AI ethics and policy planning, promotion of the Qlean Dataset service for AI training, and oversight of the Visual Library business. Appointed President & CEO in May 2024. Served as committee member for the Ministry of Economy, Trade and Industry’s study group on the “Guidebook on Generative AI Utilization for Content Production” in the same year.
Shoma Akamatsu
Tech Lead, Visual Bank Inc.
Cooperative Researcher, AIST
Graduate of Tohoku University’s School of Engineering. Conducted six years of research on quantum spintronics magnetic sensors beginning in 2018, including a visiting researcher position at MIT in 2023. Received Ph.D. in Engineering in 2024 and continued research at AIST on the social implementation of magnetic sensor technologies. Gained broad engineering experience through participation in university-based startups at Tohoku University and MIT, covering signal processing with machine learning and mathematical analysis, circuit design based on sensor elements, and physical simulation using 3D CAD. Joined Visual Bank in 2024 as an AI development engineer.
― Q. To begin, could you tell us about the role of AIST Solutions within the AIST Group?
【Mr. Osaka】
AIST, under the Ministry of Economy, Trade and Industry, is one of Japan’s largest public research institutes, with the mission of solving social issues and strengthening industrial competitiveness. To promote an open innovation ecosystem in Japanese industry, AIST Solutions was established in April 2023 as a group company.
At AIST Solutions, the role is to accelerate open innovation and create new businesses by combining proactive marketing activities with the technological assets and research resources of AIST and delivering them to companies.
About 30 years ago, Japan was often described as “Japan as Number One,” with private companies astonishing the world with their technological strength. However, after the bubble economy collapsed and as industry structures opened up, Japan’s international competitiveness declined, leading to what is often called the “lost three decades.” That said, Japan’s technological strength itself was not lost. Many of the key component technologies still used in smartphones today, such as lithium-ion batteries, CMOS sensors, OLED displays, and ceramic capacitors, originated in Japan.
Even when individual technologies excel in “time to market,” challenges remain in “time to design,” aligning with new market needs, and in “time to scale,” making early investments with future demand in mind. If Japan’s world-class technological capabilities are paired with stronger marketing strengths, a resurgence of Japanese industry will be possible.
【Mr. Mochizuki】
The activities of AIST Solutions strongly overlap with the philosophy of Visual Bank.
The company began in 2022 with the acquisition of Amana Images Inc., one of Japan’s largest stock photo agencies. The business is to operate a data library that manages rights distribution, taking visual data from rights holders and delivering it to those who need it. Just as AIST Solutions provides technology to support corporate innovation, we also aim to support innovation by providing data at the frontlines. The management philosophy of being the “invisible enabler of creativity” was born from this way of thinking.
【Mr. Osaka】
The presence of a “behind-the-scenes enabler” is extremely important when it comes to creating new design and innovation.
【Mr. Mochizuki】
I have always believed that delivering high-quality data across industries can serve as a starting point for new innovation and creativity, including AI development. That is why I strongly resonate with the vision you have described,
Mr. Osaka. It is an honor that Visual Bank has been chosen as a partner with AIST and AIST Solutions for social implementation. By combining data, technology, and marketing, we hope to make a meaningful contribution to Japan’s industry.
A One-of-a-Kind Synergy Realizing the World’s First* Image-Generation AI Model
― We heard that this joint research originated from the concept of the National Institute of Advanced Industrial Science and Technology (AIST).
【Mr. Tanaka】
Since its establishment in 2001, the National Institute of Advanced Industrial Science and Technology (AIST) has continued robotics research that began in its predecessor, the Electrotechnical Laboratory. Among its researchers, those responsible for computer vision had long been engaged in image-recognition studies.
In 2012, the advent of deep-learning-based image-recognition methods, which far outperformed traditional algorithms, triggered a major shift in the research landscape. In response, AIST established the Artificial Intelligence Research Center (AIRC) in May 2015.
Within the center, researchers focusing on image recognition shared two key questions:
How can a Japanese national research institute remain competitive in a field dominated by global Big Tech companies?
How can we prepare image data that is both safe to use and minimizes copyright or other rights-related risks?
After several years of investigation, the team discovered that if images could be generated using mathematical formulas, it would be possible to create synthetic training datasets free from copyright infringement or privacy concerns.
(Reference: Development of AI That Does Not Require Large-Scale Real-Image Collection. (available in Japanese only))
◆ What Is a Formula-Driven Model Based on Supervised Learning?
By applying fractal geometry, a universal mathematical structure characterized by self similarity in which each part resembles the whole, the researchers automatically generated image datasets for AI pre-training.
The results showed that models trained on mathematically generated images achieved recognition accuracy comparable to models trained on real image datasets.
Further analysis revealed that AI trained on fractal geometry images mainly focuses on contour features when identifying objects.
【Mr. Tanaka】
This formula driven model has already been applied in collaborative research with private companies, including autonomous driving and medical image analysis.
One example is endoscopic diagnosis of bladder cancer. Although the number of bladder cancer cases is roughly one tenth of gastric cancer, synthetic images generated through fractal geometry can supplement limited medical data for pre-training.
【Mr. Akamatsu】
So after starting with image recognition AI, the research naturally evolved toward image generation AI development.
【Mr. Tanaka】
That is correct. While image recognition focuses on identifying existing images, image generation AI allows users to freely create new visual outputs. This flexibility makes the safety of training data even more important.
At present, only a small number of generative AI models can be described as fully safe. Models that pose rights related risks are difficult to use for business applications.
If we can develop a Japan made image generation AI foundation model that minimizes rights risk, companies will be able to adopt generative AI with confidence. To achieve this, we sought partners that possess large volumes of rights cleared and safely usable visual data.
【Mr. Mochizuki】
Visual Bank Inc. has extensive experience and expertise in rights clearance, providing datasets that ensure users can safely and legally utilize data.
In addition to images and videos, we manage text, audio, and 3D data.
By combining our rights verified datasets with the technologies of AIST, we are confident that we can develop an image generation AI model that minimizes legal and ethical risks.
After examining several possible forms of collaboration, we concluded that a joint research framework would be the most effective, and therefore decided to proceed together.
【Mr. Tanaka】
From AIST’s perspective, this partnership represents an ideal synergy. Through this collaboration, we believe we can aim for the world’s first Japan made foundation model for generative AI.
We were impressed by the large scale rights cleared training datasets that Visual Bank Inc. possesses and by the company’s strong commitment to responsible data use. AIST has also assigned several of its top researchers to this joint project.
* Refers to an “image generation AI model that minimizes rights risk by combining fractal-based synthetic images with real-world images.”
(See the continuation in Part II.)
