Introducing the Ukiyo-e generation AI 'Evo-Ukiyoe' and the Ukiyo-e colorization AI 'Evo-Nishikie'
![](https://i.gzn.jp/img/2024/07/23/evo-ukiyoe/00_m.jpg)
Sakana AI, a Japanese AI company, has announced two AI models: 'Evo-Ukiyoe,' which generates ukiyo-e, one of Japan's most representative art forms, and 'Evo-Nishikie,' which colors ukiyo-e.
AI that learned from Japanese beauty: Ukiyo-e style image generation model Evo-Ukiyoe and Ukiyo-e colorization model Evo-Nishikie released
Ukiyo-e is well known around the world and is known to have influenced many great artists such as Monet and Van Gogh. However, existing image generation models have not been trained with a focus on ukiyo-e, and many people have tried to get image generation AI to output ukiyo-e, but have been unable to generate images that look like ukiyo-e.
Sakana AI's new AI models, Evo-Ukiyoe and Evo-Nishikie, are based on the company's Japanese-language-compatible image generation model 'Evo-SDXL-JP,' and are trained on 24,038 digital images of ukiyo-e prints housed at the Ritsumeikan University Art Research Center.
To develop Evo-Ukiyoe, Sakana AI first generated captions for ukiyo-e images using a technique called few-shot prompting , which teaches the model by giving concrete examples. It then trained the base model, Evo-SDXL-JP, using pairs of image data and captions as a learning dataset, and fine-tuned it using LoRA.
Evo-Ukiyoe, which was created in this way, can generate images that are similar to ukiyo-e by using prompts such as cherry blossoms, Mt. Fuji, kimonos, and birds that are often featured in ukiyo-e. On the other hand, if prompts are used for things that did not exist in the Edo period, such as computers or hamburgers, the images generated may not look like ukiyo-e.
![](https://i.gzn.jp/img/2024/07/23/evo-ukiyoe/1_m.jpg)
For example, if you use the prompt 'There are plants and flowers. Butterflies are flying. Top quality spoke ukiyo-e.', the result will be as follows. Note that the phrase 'spoke ukiyo-e' is a special token used during learning, that is, a token that has been trained differently from normal words in order to give it a specific function.
![](https://i.gzn.jp/img/2024/07/23/evo-ukiyoe/3_m.jpg)
If you use the prompt 'A crane is standing in the garden. It's snowing. The highest quality Japanese woodblock print.' you get something like this.
![](https://i.gzn.jp/img/2024/07/23/evo-ukiyoe/4_m.jpg)
And below is the output from the prompt 'A cat wearing a kimono drinking tea in a garden. A top quality ukiyo-e print.'
![](https://i.gzn.jp/img/2024/07/23/evo-ukiyoe/5_m.jpg)
Another generative AI, Evo-Nishikie, is based on Evo-Ukiyoe and is a model that uses condition images created by removing noise such as insect holes and blemishes from ukiyo-e images as a training dataset. It can turn monochrome ukiyo-e such as sumizuri-e into multi-colored ukiyo-e, also known as nishikie.
Below is a monochrome ukiyo-e print (left) from the Ritsumeikan University Art Research Center collection, which has been colorized using Evo-Nishikie (center and right).
Ryo Akama, director of Ritsumeikan University's Art Research Center, said, 'There are contemporary artists all over the world who create works based on ukiyo-e. The ukiyo-e created by Evo-Ukiyoe will not only inspire these artists' creative desire, but will also provide an opportunity for general enthusiasts to explore more deeply the meaning of authentic ukiyo-e as classics.'
In addition, one of the developers of Evo-Ukiyoe and Evo-Nishikie, Sakana AI's Karanuwat Tarin, is a person who has developed the model 'RURI' used in the AI cursive character recognition application 'Mio'.
I tried using the AI cursive character recognition app 'miwo' - GIGAZINE
![](https://i.gzn.jp/img/2021/08/31/miwo/00_m.jpg)
Related Posts: