Description

Hitoshi Ashinano is a Japanese manga artist known for his contemplative, nostalgic, and atmospheric storytelling. His most famous work is Yokohama Kaidashi Kikō, with which he debuted in 1994. It is a slice-of-life science fiction manga set in a serene, post-apocalyptic world, serialized in Monthly Afternoon until 2006. The series won the 2007 Seiun Award for Best Science Fiction Manga and was adapted into two OVA series.

Before his debut, Ashinano worked as an assistant to Kōsuke Fujishima, the creator of Oh My Goddess!. Another notable work is PositioN (1999-2001), a short manga exploring magical realism in everyday life. His later works include Kabu no Isaki, which started as a one-shot in 2007 before becoming a full series.

Ashinano's works are characterized by minimal dialogue, strong female protagonists, and a deeply immersive atmosphere, often evoking a sense of tranquility and wonder.

Usage

Trigger words are Hitoshi Ashinano style.

All images that I post contain workflow metadata for ComfyUI. But just for reference, I used:

shift: 6.0
steps: 40
cfg: 4.0
sampler: gradient_estimation
scheduler: sgm_uniform

Training

I used 113 high-res images from "Yokohama Kaidashi Kikou".

For captioning I used JoyCaption AlphaTwo in "descriptive/long" prompting mode. All captions were prefixed with the phrase "You are an assistant designed to generate high-quality images based on user prompts. <Prompt Start> Hitoshi Ashinano style. "

Training was done with ai-toolkit on an RTX 3090, Windows 11, with 64 GB RAM. I planned to train until 20000 steps, but stopped at 10000, because there were already plenty of good checkpoints to choose from, and the average loss had stopped decreasing. And I could not wait any longer to test LoRA in production :) After testing various prompts, I chose the checkpoint at 9000 steps.

Training hyperparameters were default, except:

network.linear: 32
network.linear_alpha: 32
datasets.caption_dropout_rate: 0.1
optimizer: lion
optimizer_params.betas: [0.95, 0.99]
optimizer_params.weight_decay: 0.05
optimizer_params.lr: 2e-5
optimizer_params.noise_offset: 0.1
optimizer_params.lr_scheduler: constant