While the models only support prompting in Japanese and are in the very early stages, Sakana hopes the work to teach AI traditional “Japanese beauty” will spread the appeal of the country’s culture worldwide and find applications in education and new ways of enjoying classical literature.

Currently, the company is providing both models and the associated code to get started on Hugging Face. The Python script included in the repository and LoRA weights are available under the Apache 2.0 license.

“This model is provided for research and development purposes only and should be considered as an experimental prototype. It is not intended for commercial use or deployment in mission-critical environments. Use of this model is at the user’s own risk, and its performance and outcomes are not guaranteed,” the company notes on Hugging Face.

So, far Sakana AI has raised $30 million in funding from multiple investors, including by Lux Capital, which has invested in pioneering AI companies like Hugging Face, and also Khosla Ventures, known for investing in OpenAI way back in 2019. Now, with the release of the two image-generation models, Sakana wants to bring the historic artwork back into popular culture. The first one – Evo-Ukiyoe – is a text-to-image offering that generates images closely resembling ukiyo-e, especially when prompted with text inputs describing elements commonly found in ukiyo-e art such as cherry blossoms, kimono or birds. It can even generate ukiyo-e-style art with things that did not exist back then, like a hamburger or laptop, but the company points out that sometimes the results may veer off track — not resembling ukiyo-e at all.

The model is based on Evo-SDXL-JP, which Sakana developed using its novel evolutionary model merging technique on top of Stability AI’s SDXL and other open diffusion models. The company said it used LoRA (Low-Rank Adaptation) to fine-tune Evo-SDXL-JP 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.

“We curated this data with a wide range of subjects, covering including whole art and face-centered ones, from the digital images of ukiyo-e in the ARC collection. We also focused on multi-colored nishiki-e with beautiful colors while considering diversity,” the company wrote in a blog post.

The second model, Evo-Nishikie, is an image-to-image offering that colorizes monochrome Ukiyo-e prints. Sakana says it can add color to historical book illustrations that were printed in one color of ink or give entirely new looks to existing multi-colored Nishikie prints. All the user would have to do is provide the source image and maybe pair it with a set of instructions describing the elements to be colored. Sakana said it brought this model to life by performing ControlNet training on Evo-Ukiyoe, using fixed prompts and condition images.