Frontier of Embodiment Informatics: ICT and Robotics, Top Global University ProjectWaseda University

I conducted research on hierarchical image generation using diffusion models, one of the leading techniques in image synthesis. A diffusion model is a generative method that estimates a data distribution by gradually transforming real image data into a random state and then defining the reverse process to reconstruct meaningful structures. Originally developed from Langevin dynamics in physics to describe molecular motion, diffusion models have a clear physical and mathematical foundation.

In practice, however, diffusion models are not implemented solely on the basis of theory. Instead, they incorporate various refinements in network architectures and inference algorithms. As a result, the theoretical ideal and the practical optimal solution do not always coincide, and real-world implementations often rely on heuristic adjustments. This demonstrates that there remains room for bridging the gap between theory and practice.

Based on this challenge, I drew inspiration from the process by which humans create drawings. Typically, one begins by sketching the overall contours, then progressively adds shapes and colors to different parts, and finally fills in the details. By starting with coarse information and gradually adding finer details, each step clarifies what should be drawn next, resulting in a well-structured composition. I applied this principle to diffusion models, designing the generative process in a coarse-to-fine manner. This approach aimed to improve the separability and controllability of the information added at each stage, thereby enabling more intuitive and controllable image generation.

Throughout the course of this research, I carried out the entire process from reviewing prior work to implementing the proposed method, gaining significant insights into research design and evaluation methodologies. Furthermore, I developed a visualization tool that has proven useful for understanding generative processes and comparing models, and it has been utilized in the host laboratory abroad. Although I was not able to bring the research to full completion during the exchange period, I am continuing to refine the work after returning to Japan, with the goal of submitting the results to a top international conference.

International Exchange

In the laboratory, both English and Finnish were spoken. Through research discussions, casual conversations, and various other interactions, I was able to develop my ability to explain ideas in English, and I gained greater confidence in speaking English compared to before my stay. I also studied Finnish little by little, which gave me an opportunity to learn about the country’s history and cultural values. In addition, I lived in a shared apartment with students from across Europe. By traveling together and participating in various activities, I was able to deepen my understanding of the different values and perspectives each of us held.

Dinner with laboratory members

Residential Environment

I stayed in a shared apartment located in Espoonlahti (which means “Bay of Espoo” in Finnish), about 15 minutes by metro from Aalto University. The area was surrounded by forests, providing a very quiet and peaceful environment. Because Finland is situated at a high latitude, when I first arrived in February the sun rose around 8 a.m. and set around 5 p.m. In contrast, as the summer solstice approached, daylight remained even past 11 p.m. This allowed me to experience unique ways of spending time, such as going out for drinks with fellow students after research or enjoying the sauna—experiences that would be difficult to have in Japan.

Peaceful time spent at a gulf near my apartment

Surrounding Environment

Approximately 74% of Finland’s land area is covered by forests, and there are around 190,000 lakes scattered across the country. With such abundant nature, it is common for local people to spend time in the forest after work, either conversing with friends or enjoying solitude. I too often visited the forest when I felt tired from research, taking time to relax and clear my mind. From the laboratory, the center of Helsinki was easily accessible by metro or bus, making shopping and sightseeing convenient. Around my apartment, in addition to forests and a harbor, there was also a large shopping complex, so I had no difficulty purchasing daily necessities and groceries. Furthermore, Estonia, located across the Gulf, could be reached by ferry in about two hours, which allowed me to visit for sightseeing as well.

Summer in Lapland

Tallinn, the Capital of Estonia, Retaining Traces of the Medieval Era

In Finland, coffee breaks are guaranteed by labor law, and multiple coffee machines are installed on campus. As a result, it was common for laboratory members to gather with a cup of coffee in hand and engage in discussions, which left a strong impression on me. The laboratory also had extra chairs, and professors as well as students would bring them over to my desk or to other students’ desks to hold consultations and discussions. I felt that the constant readiness to explain one’s research fostered an atmosphere that encouraged lively academic exchange.

On the other hand, winters in Finland are extremely harsh, and perhaps for that reason, people eagerly anticipate the arrival of spring. In particular, May 1st is celebrated as Vappu, a holiday marked by traditional sweets and drinks. Furthermore, on the day of the summer solstice, the Midsummer Festival is held, during which it is customary to spend time with family at a summer cottage. During this season, I enjoyed the sunset at a beach near Helsinki, where I could truly experience the changing of the seasons.

Summer solstice in Finland