Sakana AI brings Japan's historic ukiyo-e to life with AI

Sakana AI has open-sourced Evo-Ukiyoe and Evo-Nishikie, two image generation models designed to create traditional Japanese ukiyo-e woodblock prints from text and colorize centuries-old monochrome book illustrations.

The models split the creative workflow down the middle. Evo-Ukiyoe is a text-to-image generator tuned specifically for classical floating world aesthetics. Feed it classic elements like cherry blossoms, kimonos, or birds, and it delivers authentic Edo-period styling. You can also throw modern concepts at it, like a burger or a laptop, though the creators note that anachronistic prompts can sometimes go wildly off-track.

Evo-Nishikie handles image-to-image colorization. Drop in a single-color historical woodcut illustration along with descriptive Japanese prompt instructions, and it transforms the bare linework into a vibrant multi-colored nishiki-e print. It can also give existing vintage color prints an entirely fresh palette.

How it was built

Neither model was trained from scratch. Instead, Sakana AI built upon its Evo-SDXL-JP foundation, a Japanese-centric base developed by merging Stability AI's SDXL with open-source diffusion models via evolutionary model merging.

The real secret sauce lies in the training data. Sakana AI partnered with the Art Research Center at Ritsumeikan University in Kyoto, securing access to a curated dataset of over 24,000 meticulously captioned ukiyo-e artworks. The collection covers diverse historical themes, full landscapes, and facial portraits.

Technically, Evo-Ukiyoe was trained using LoRA fine-tuning on that dataset. For the colorizer Evo-Nishikie, the team applied ControlNet training on top of Evo-Ukiyoe using paired condition images and fixed text prompts.

Requirements and constraints

Both models, inference code, and LoRA weights are openly available on Hugging Face under an Apache 2.0 license.

There are a couple of practical hurdles to keep in mind. The models currently only understand Japanese text prompts, meaning English inputs will not produce the intended results.

Sakana AI also explicitly labels these releases as experimental research prototypes. They are intended for academic and non-commercial exploration rather than mission-critical deployment.

Why it matters

Broad generic diffusion models can imitate ukiyo-e surface textures, but they consistently bungle historical composition, accurate pigmentation, and woodblock grain. For digital humanists, educators, and indie illustrators, having an open-source model trained on museum-grade archives turns tedious archival restoration and heritage rendering into a one-click workflow.

It is also a textbook lesson in smart AI specialization. Instead of burning millions attempting to outmuscle global foundation giants, fine-tuning an open base with pristine academic collections allows specialized teams to build deep, culturally rich tools that mass-market models simply cannot match.