Sakana AI, a Tokyo-based startup, has made waves in the AI industry with its innovative approach to developing high-performing models. The company, founded by former Google employees, has recently announced two new image-generation models: Evo-Ukiyoe and Evo-Nishikie. These models are designed to generate images from text and image prompts, with a unique focus on Japan’s popular historic art form ukiyo-e. Ukiyo-e, which emerged in the 17th to 19th centuries, was a genre of art that depicted scenes from historical events, landscapes, and sumo wrestlers. Sakana AI aims to bring this art form back into modern content consumption using the power of AI.
The release of these two image-generation models is part of a growing trend in the AI industry: localization. Companies in countries like South Korea, India, and China have been building models tailored to their respective cultures and dialects. This localization effort recognizes the importance of incorporating cultural references and nuances into AI models to better serve specific communities.
Evo-Ukiyoe, the first model introduced by Sakana AI, is a text-to-image offering that generates images resembling ukiyo-e art. When prompted with text inputs describing elements commonly found in ukiyo-e, such as cherry blossoms or kimono, the model can create ukiyo-e-style art. Interestingly, it can even generate ukiyo-e-style art with modern elements like hamburgers or laptops. However, the company acknowledges that sometimes the results may not resemble ukiyo-e at all.
The model is based on Evo-SDXL-JP, which Sakana developed using its novel evolutionary model merging technique. To fine-tune the model, Sakana used LoRA (Low-Rank Adaptation) on a dataset of over 24,000 carefully-captioned ukiyo-e artworks acquired through a partnership with the Art Research Center (ARC) of Ritsumeikan University in Kyoto. The company curated the dataset to ensure a wide range of subjects and diversity in the art.
The second model, Evo-Nishikie, is an image-to-image offering that colorizes monochrome ukiyo-e prints. It can add color to historical book illustrations that were originally printed in one color of ink or give entirely new looks to existing multi-colored Nishikie prints. Users simply need to provide the source image and instructions describing the elements to be colored.
Sakana AI hopes that by teaching AI about traditional “Japanese beauty,” it can spread the appeal of Japan’s culture worldwide and find applications in education and new ways of enjoying classical literature. While the models currently only support prompting in Japanese and are in the early stages, Sakana is providing both models and associated code on Hugging Face for research and development purposes.
It’s worth noting that these models are experimental prototypes and not intended for commercial use or deployment in mission-critical environments. Use of the models comes with its own risks, and their performance and outcomes are not guaranteed. Sakana AI has raised $30 million in funding from investors such as Lux Capital and Khosla Ventures, which demonstrates the confidence in the company’s vision and potential impact in the AI industry.
In conclusion, Sakana AI’s new image-generation models, Evo-Ukiyoe and Evo-Nishikie, offer a unique approach to generating images inspired by ukiyo-e art. By incorporating cultural references and leveraging AI technology, Sakana AI aims to revive the popularity of this historic art form and bring it into modern content consumption. The company’s localization efforts reflect a broader trend in the AI industry, where companies are building models tailored to specific cultures and dialects. With continued research and development, Sakana AI hopes to further explore the applications of AI in promoting traditional Japanese beauty and enhancing education and literature experiences.