Six urban scenes generated using Stable Diffusion XL and a LoRA adapter trained on the woodblock prints of Katsushika Hokusai.
01 / Idea
How will a familiar urban scene change if depicted using the techniques of Japanese color woodblock printing?
In this project, I explored transplanting Hokusai’s visual language onto everyday life scenes: a walk along the embankment, a crossing of a river, a courtyard, a construction site, a winter park, and an evening stop.
I chose Hokusai for the combination of expressive outline, large fields of color, and detailed drawing within them. Initially, I aimed to render contemporary urban scenes in this chosen graphic style. During the generation process, I discovered that the model transfers characteristic architecture, clothing, and landscape motifs along with the color and line. The final series demonstrates how the urban subject transforms under the influence of the training dataset.
02 / Source Images
To train the model, I collected 24 reproductions from different engravings in the «Thirty-Six Views of Mount Fuji» series. I limited the selection to one series to maintain the consistency of the graphic style and included various types of scenes: water, mountains, settlements, bridges, and daily labor.
Image reproductions sourced from The Metropolitan Museum of Art collection. For the selected images, museum catalog cards and Public Domain markings have been preserved. The museum distributes such reproductions under an Open Access framework with a CC0 designation.[Image Usage Terms]
Below are six examples from the original selection. These display full museum reproductions; for the study, I used square fragments.
A Big Wave in Kanaga
Edziri in Suruga Province
The Waterwheel in Ondene
In the mountains of Totomi Province
Morning after a snowfall in Koishikawa
Nihonbashi in Edo
03 / Training and Education
Horizontal reproductions were cropped and resized to 1024 × 1024 pixel squares without stretching or generative inpainting. This transformation allowed the dataset to conform to a uniform format, though part of the original composition fell outside the frame.
For each image, I prepared a caption describing its content, along with the general marker hksai24 woodblock print. This connects the images in the collection to a single identifier in the model prompts.The training took place in Google Colab on a Tesla T4. I used SDXL Base 1.0 and the DreamBooth method with LoRA: the UNet adapter was trained, while the text encoders remained frozen. The prepared files were 1024 × 1024, but the actual training was performed at a resolution of 512 × 512.
The journal confirms the achievement of 800 steps and the retention of final weights. Checkpoints were logged every 200 steps. The main run, from start to completion, took approximately 1 hour and 26 minutes. After resetting the session, I used the weights, configuration, and journal saved on Google Drive; retraining was not required to generate results.
04 / Final Series
I gathered six 1024×1024 pixel images. They are unified by a cream background, shades of blue and green, subtle ochre and red accents, dark outlines, and a simulated paper texture. Each scene, however, possesses its own compositional rhythm.
1 — Embankment
The diagonal embankment connects a dense group of passersby with the open space of the river. The umbrellas form a repeating pattern, and the elongated boats guide movement along the shore. Instead of a contemporary urban environment, the model proposes historical architecture and attire.
2 — Crossing
A large arch of a metal bridge dominates the upper portion of the composition; below it, boats and riverbanks are situated. The red train cars stand out against the light background. The transportation narrative is maintained, yet it appears fantastical: the train is fragmented, and the bridge structure and movement across it are depicted symbolically. A familiar silhouette of Mount Fuji, found in the dataset, appears on the horizon.
3 — Courtyard
Vertical walls, sloped roofs, and the silhouette of a pine tree frame a small courtyard space. This piece is calmer than the others in its density of action. However, the request for a courtyard between apartment buildings has been transformed into wooden construction; two dogs appear instead of one. It is particularly noticeable here that the model carries over both the style of rendering and the subject motifs.
