Sakana AI releases experimental models for ukiyo-e-style generation and print colorization

ThatPainter is reader-supported. When you buy through links on our site, we may earn an affiliate commission. Learn More

Sakana AI announced two specialized image models on July 21, 2024: Evo-Ukiyoe-v1 generates new ukiyo-e-style images from Japanese text prompts, while Evo-Nishikie-v1 colorizes monochrome or line-based woodblock-print imagery. Both were trained from Sakana AI’s Japanese-language Evo-SDXL-JP foundation using 24,038 digitized ukiyo-e images selected from works held by Ritsumeikan University’s Art Research Center.

The release is significant for culturally focused image generation, but it is not a new consumer art app or a replacement for traditional printmaking. Sakana’s public v1 model cards describe the models as experimental and intended for research and education, not commercial production or mission-critical use.

#	Preview	Product	Price
1		Japanese Woodblock Prints. 45th Ed. (45th Edition)	$27.46	Buy on Amazon
2		Japanese Woodblock Prints (Basic Art)	$15.43	Buy on Amazon
3		Japanese Woodblock Print Workshop: A Modern Guide to the Ancient Art of Mokuhanga	$18.79	Buy on Amazon
4		Pictures of the Floating World: An Introduction to Japanese Prints	$28.00	Buy on Amazon
5		Japanese Woodblock Prints: Artists, Publishers and Masterworks: 1680 - 1900	$28.69	Buy on Amazon

Two models, two different jobs

Reports describing Sakana AI’s release as a single “ukiyo-e AI” miss its most important distinction. Evo-Ukiyoe and Evo-Nishikie accept different inputs and are designed for different creative tasks.

In practical terms, Evo-Ukiyoe is suited to prompts about landscapes, figures, clothing, and other scenes associated with ukiyo-e. Evo-Nishikie starts with an existing image. It can explore how a monochrome illustration might look as a multicolor nishiki-e print, including images from historical books.

#1 Best Overall

Sale

Japanese Woodblock Prints. 45th Ed. (45th Edition)

$27.46

Buy on Amazon

That means Evo-Nishikie is not simply a second text-to-image model. Its defining feature is image conditioning: the source image helps determine the composition while the prompt guides the generated color treatment and interpretation.

Why build a specialized ukiyo-e model?

Sakana AI’s stated argument is that generic image models often treat “ukiyo-e” as a broad label for Japanese-looking illustration. They may produce an attractive image with familiar motifs without consistently capturing features associated with historical woodblock prints: strong contour lines, flat areas of color, distinctive compositions, recurring subject matter, and the visual logic of printed images.

A model trained and tuned around ukiyo-e imagery can make those characteristics more central to generation. Sakana AI presents the project as a way to support cultural and historical education, encourage interest in ukiyo-e and Japanese culture, and explore Japan-specific AI development rather than relying only on a generic model’s interpretation of a cultural term.

That is a specialization claim, not proof that the models outperform every current general-purpose image generator. The 24,038-image figure describes the training material; it does not establish equal coverage of every artist, school, period, subject, or region, nor does it provide an independent quality benchmark.

What was in the training set?

According to Sakana AI’s announcement, the models used 24,038 digitized images of ukiyo-e works selected from the collection of Ritsumeikan University’s Art Research Center. The selection was made with the center and emphasized attractive color palettes and subject diversity.

Rank #2

Sale

Japanese Woodblock Prints (Basic Art)

$15.43

Buy on Amazon

The material included full images as well as face-centered crops. That detail matters: crops can help a model learn facial and figure-related visual information, but the dataset should not be interpreted as a complete survey of ukiyo-e history. The number also refers to selected digital images, not necessarily 24,038 unique physical prints.

The models were built on Evo-SDXL-JP, Sakana AI’s Japanese-language image-generation foundation. Sakana describes Evo-SDXL-JP as the product of its “evolutionary model merging” approach, after which the ukiyo-e models were specialized using the curated print imagery.

Does it create authentic ukiyo-e?

No—not in the historical or physical sense. The defensible description is ukiyo-e-style digital generation.

A generated image does not reproduce the processes that define an actual woodblock print: carving separate blocks, preparing pigments, selecting paper, aligning the blocks through registration, and making successive impressions by hand. It also does not guarantee that the clothing, architecture, tools, social roles, geography, or iconography in a scene are historically correct.

The models may produce imagery with visual features associated with ukiyo-e, and Sakana says common subjects such as landscapes and people in kimono can be generated with a closer resemblance to ukiyo-e than they would be by a generic model. That remains a company description rather than an independently measured comparison.

Rank #3

Sale

Japanese Woodblock Print Workshop: A Modern Guide to the Ancient Art of Mokuhanga

$18.79

Buy on Amazon

There is a second risk: specialization can create a coherent “house style” while flattening important differences among artists, schools, periods, and traditions. A result that looks convincingly Edo-period at a glance may combine elements that never belonged together historically.

Evo-Nishikie is color interpretation, not automatic restoration

Evo-Nishikie can be useful for educators, researchers, artists, and museums exploring possible color treatments for monochrome illustrations. Sakana’s examples include work based on Ehon Tamakatsura, a classical book published in 1736.

But a generated color image should not be presented as proof of the original colors unless those colors are independently documented. The model is making an informed visual interpretation from its input and prompt. It can produce plausible colors that are historically unsupported, especially when the source image lacks enough evidence about pigments, fading, or the original printing process.

For archival or museum work, retain the original scan beside the generated version, label the AI alteration clearly, and describe the output as an interpretation rather than a recovered original.

How to try the models

The public v1 repositories and demo are available through Hugging Face, although hosted demos, dependencies, and instructions can change:

Rank #4

Sale

Pictures of the Floating World: An Introduction to Japanese Prints

$28.00

Buy on Amazon

• Evo-Ukiyoe-v1 model repository
• Evo-Nishikie-v1 model repository
• Evo-Ukiyoe demo

A cautious workflow is:

1. Open the current model card and read its license, restrictions, hardware requirements, and installation notes.
2. Use the hosted demo where available, or prepare a local Python environment and install Git LFS before downloading large model files.
3. For Evo-Ukiyoe, write a Japanese prompt describing the subject, setting, composition, and print characteristics.
4. For Evo-Nishikie, provide a clean monochrome or line-processed source image and a concise Japanese description of its contents.
5. Generate several variations and inspect anatomy, faces, hands, lettering, symbols, historical details, and color choices.
6. Document the model version, prompt, source image, and any edits if the result will be used in an educational, institutional, or published project.

The model repository shows a Git-based download beginning with:

git clone https://huggingface.co/SakanaAI/Evo-Ukiyoe-v1

That command alone is not a complete installation recipe. Python, CUDA, PyTorch, diffusion-library, GPU-memory, and dependency requirements should be taken from the live model card rather than copied from an older guide.

What to expect from the output

Japanese prompts help, but do not solve everything

The models were designed around Japanese-language prompting, so prompts should describe concrete visual goals rather than relying on a vague phrase such as “Japanese art.” You might specify a woodblock-print composition, limited flat colors, bold contour lines, an Edo-period landscape, traditional framing, or a particular subject category.

Prompt wording cannot guarantee historical authenticity. It also does not guarantee accurate Japanese lettering inside the generated image. For posters, labels, book covers, and other text-heavy work, typeset the wording separately.

Figures and details may need correction

Faces, hands, repeated figures, costumes, tools, and architectural details can require multiple generations and manual cleanup, as they do with many diffusion-based systems. Generated outputs should be reviewed rather than treated as finished historical illustrations.

Best Value

Sale

Japanese Woodblock Prints: Artists, Publishers and Masterworks: 1680 - 1900

• Used Book in Good Condition

$28.69

Buy on Amazon

Color may be visually persuasive but wrong

If Evo-Nishikie produces implausible colors, try a cleaner source image, provide a more specific description, and compare multiple seeds or outputs. Keep the result labeled as an AI-generated interpretation.

Commercial use requires separate legal checks

Publicly available weights and a public demo do not automatically make the outputs commercially cleared. The Evo-Ukiyoe-v1 model card describes the model as experimental, for research and educational purposes, and not intended for commercial use or mission-critical deployment. Check the current terms for both models before using them in merchandise, advertising, client work, publishing, or a paid product.

Several rights questions remain distinct:

• The rights and permissions associated with the digitized source images.
• The license governing the model weights and code.
• Any rights attached to generated outputs under the applicable jurisdiction and service terms.
• Rights in recognizable people, protected characters, logos, or institutional collections.
• Permission to publish altered archival or museum material.

For commercial client work, a maintained image service with explicit commercial-use terms may be easier to evaluate, but that is a licensing and workflow comparison—not evidence that services such as Adobe Firefly or Leonardo.Ai generate better ukiyo-e. Neither should be treated as a substitute for the specialized research purpose of Sakana’s models without comparative testing.

What the release is—and is not

Sakana AI’s July 2024 announcement is best understood as a research release of two culturally specialized image models:

• Evo-Ukiyoe creates new ukiyo-e-style images from Japanese text prompts.
• Evo-Nishikie applies image-conditioned generation to colorize or reinterpret monochrome print imagery.
• Both use a Japanese-language foundation and a curated ukiyo-e dataset from Ritsumeikan University’s Art Research Center.
• Neither guarantees historical accuracy, authentic printmaking, original pigment recovery, or commercial readiness.

A later Tokyo Metropolitan Government profile published in 2025 referred to Evo-Ukiyoe v2 as under development. That reference does not establish a public v2 release, public weights, a license, or a production service. The publicly documented v1 release remains the appropriate basis for evaluating what readers can use.

Quick Recap

SaleBestseller No. 1

Japanese Woodblock Prints. 45th Ed. (45th Edition)

$27.46

SaleBestseller No. 2

Japanese Woodblock Prints (Basic Art)

$15.43

SaleBestseller No. 3

Japanese Woodblock Print Workshop: A Modern Guide to the Ancient Art of Mokuhanga

$18.79

SaleBestseller No. 4

Pictures of the Floating World: An Introduction to Japanese Prints

$28.00

SaleBestseller No. 5

Japanese Woodblock Prints: Artists, Publishers and Masterworks: 1680 - 1900

Used Book in Good Condition

$28.69

ThatPainter Team

Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.