AI That Learned Japanese Beauty: Sakana AI Releases Evo-Ukiyoe and Evo-Nishikie Models
Sakana AI
Sakana AI released two new image models trained on real ukiyo-e woodblock prints from a Japanese university archive.
Based on reporting by Sakana AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Sakana AI just dropped two models that feel like a genuinely different use of generative AI than the usual meme-image churn. Evo-Ukiyoe turns Japanese text prompts into images in the style of classic ukiyo-e prints, while its sibling Evo-Nishikie takes old single-color sumizuri-e illustrations and colorizes them into the vibrant, multi-block nishiki-e style that made prints like Hokusai's Great Wave famous.
The interesting part is the data. Instead of scraping the web for anything tagged "ukiyoe," which tends to produce vaguely Japanese-looking illustrations rather than actual woodblock aesthetics, Sakana AI partnered with Ritsumeikan University's Art Research Center. Professor Ryo Akama's team helped select 24,038 high-resolution images from the ARC's ukiyo-e archive, prioritizing color richness and subject diversity. That curated dataset became the foundation for fine-tuning Evo-SDXL-JP, Sakana's earlier Japanese-language image model built through evolutionary model merging.
Technically, the two models take different paths. Evo-Ukiyoe is a straightforward text-to-image system, fine-tuned with LoRA on image-caption pairs generated by a multimodal model and then manually cleaned up to fit SDXL's 77-token prompt limit. Evo-Nishikie works more like a colorizing pipeline: it uses ControlNet trained on lineart extracted from the original prints, with OpenCV cleaning up centuries of insect damage, staining, and fading before feeding it a fixed, simple prompt like
My take — AI-written commentary, not fact-checked reporting
I like that this isn't another chatbot wrapper or a Midjourney clone chasing photorealism, it's a niche, culturally specific tool built on a genuinely curated dataset, which is exactly the kind of AI application that gets underrated because it doesn't scale to a billion users. Open-weighting it on HuggingFace for research and education is the right call too, since the actual value here is scholarly and cultural, not commercial, and locking it behind an API would have missed the point entirely.
Read more about this at: Sakana AI
