What this guide covers
In November 2026 we replaced four of EGAKU's anime styles (ukiyo-e, ink wash, Art Nouveau and Impressionism) with style models (LoRAs) we trained ourselves, using only public-domain and CC0 images. This guide is the exact procedure.
A LoRA quietly trained on other people's artwork is easy to make and easy to share, but it is hard to justify for commercial use. How good can a style model be when every training image has a clear rights status? Where did we get stuck, and how did we fix it? We wrote it down in a reproducible form: the actual settings, how we chose the images, and how we judged the results.


Scope and limits (up front)
- This is about style LoRAs for an SDXL-based anime base model. It does not carry over as-is to photographic styles or to video models.
- What the model learns is the look (line, colour, texture), not the ability to call up a specific artist by name. We remove artist names from the captions.
- We trained on a rented cloud GPU (24 GB or more). A local GPU works too, just more slowly.
- Public-domain art has blind spots. Our ukiyo-e set has almost no cats, so asking for a cat produced a person. Training settings do not fix that.
The overall flow
- Choose the images: from a collection of 12.4 million public-domain and CC0 images, pick ones that are close to the style and well made, using only the metadata and image features, without downloading the images
- Build the training set: download and resize, clean up the captions (remove artist names, cap images per artist)
- Train: 1,600 steps per style, keeping a checkpoint every 400 steps
- Choose: decide with a table of steps × strength, same prompts and same seed
The paid part gives the concrete values and the reasons for each stage, where we got stuck and how we fixed it, and the checkpoint and strength we shipped.
以下為付費內容
付費部分包含
- 1. Choosing images without downloading them
- 2. Building the training set (this decides the result)
- 3. Training settings (every value, and why)
- 4. Four environment problems we hit
- 5. Judging the result (a steps × strength table)
- 6. The checkpoint and strength we shipped
- 7. How cost works, and how to keep it down
一次性購買,立即可讀,並保存在你的帳戶中。
正在確認購買…