Tokaido 2.0

«Tokaido 2.0”—a series depicting a contemporary journey through Japan, executed in the style of Utagawa Hiroshige’s landscape prints. Electric trains, buses, and bicycles appear within a space constructed from flat fields of color, fine contours, and smooth gradations of sky and water.

The project’s goal is to train a LoRA adapter for the «Silent Keepers» series. Seven characters are connected to the key, a letter, a lantern, a cup, a mirror, seeds, and a clock. The final still life, featuring an empty frame, leaves space for the absent eighth keeper. The objects become carriers of a personal history: access, inscription, light, ritual, recognition, growth, time, and absence.

Thirty-two reproductions from the «53 Stations of the Tōkaidō» series were used for this study. The images were provided by The Metropolitan Museum of Art and carry individual CC0 licenses, which are recorded in the dataset manifest.

Stable Diffusion XL 1.0 was used with the DreamBooth LoRA method. Additional UNet weights were trained, while the text encoders remained frozen.

Key Parameters: GPU: Tesla T4; 800 steps, rank 16; learning rate: 0.0001; batch size: 1; training resolution: 512×512; fp16, gradient checkpointing, and 8-bit Adam.

800 optimization steps took 22 minutes and 23 seconds. Training completed with a return code of 0; weights were saved. 30 steps and a 1024×1024 resolution were used for generation. Subsequent retouching and generative inpainting were not applied.

The series consists of six 1024×1024 pixel images:

1. «City Bridge"—a suspension bridge against the backdrop of a modern city.
2. «Coastline"—a train moving along the sea.
3. «New Crossing"—a boat, cyclists, and a view of Mt. Fuji.
4. «Mountain Pass"—a bus surrounded by steep slopes.
5. «Snowy Stop"—passengers in a snowy landscape.
6. «Rainy Alley"—cyclists on a wet city street.

The images are united by a muted blue palette, a warm paper texture, fine dark outlines, and flat fills. These characteristics link contemporary subjects to the visual language of historical xylography.

The compositions vary in rhythm. The verticals of the bridge create a dense urban scene, the electric train continues the diagonal of the shore, and the crossing is framed around a calm water surface. The mountain pass emphasizes the scale of the rocks. The snowy stop becomes a quiet pause, and the rainy alley concludes the journey with the movement of people.

In the initial results with a LoRA strength of 0.8, some modern elements disappeared. The prompts for the bridge, electric train, and bus were refined, and variants were created using a LoRA strength of 0.55. These better conveyed the intended narratives and made it into the final selection.

Additionally, six control images from the base SDXL model were obtained. A comparison of the rainy alley, using the same prompt and seed, showed that the base model already conveys the intended stylization, while the LoRA enhances the paper texture and alters the figures' details.

Limitations remain: the conditional geometry of the bridge and road, distortions in individual figures and bicycles, and pseudo-signatures. The series represents artistic fantasy, not a documentary depiction of the route.

Stable Diffusion XL 1.0 with a trained LoRA adapter was used to create the images.

ChatGPT was used for concept development, code preparation, and prompting.

No other image generators were utilized. The source dataset consists of museum reproductions.