ceres, adult woman, long straight white hair, parted bangs, long sidelocks, blue eyes, medium breasts, mature appearance

CFG Scale: 4
Steps: 30

These are the settings that produced relatively stable results in my own testing.
These are not absolute or definitive settings, so more experienced users are welcome to adjust them freely according to their own setup and preferences.

This version performs most consistently in monochrome. Full-color generation is possible, but the colors and shading may sometimes look less natural. I plan to improve this in a future update.

First of all, I would like to sincerely thank 00X09901 for their help.

After I started using Civitai, I wanted to learn how to train LoRA models entirely on my own, without relying on other people. I studied the process by myself and made many training attempts while repeatedly rebuilding datasets and changing settings, but I was unable to produce a model that I was truly satisfied with.

After struggling through many failed attempts, I eventually asked 00X09901 for advice. They gave me a crucial hint and valuable guidance about the actual training process that helped me understand what I had been missing.

I then studied additional LoRA training resources and guides from several different sites, reorganized my dataset, tags, and training settings, and tried again.

Today was the first time I genuinely felt, “This model is finally working properly.”

Unlike my previous attempts, the character remained much more consistent with the prompts. I tested the model more than twenty times using different poses, camera angles, compositions, outfits, and aspect ratios, and the overall results remained surprisingly stable. For the first time, the different parts of the training process felt like gears finally fitting together and working as intended.

The generation settings that have produced the most stable results in my testing so far are:

CFG Scale: 4
Steps: 30

General prompts have worked consistently across different poses, angles, and outfits. I also tested adult-oriented prompts, and the model performs particularly well with monochrome generations.

The main limitation of the current version is full-color generation. Since most of the original training material came from black-and-white manga, monochrome images are highly consistent, while full-color generations can sometimes retain monochrome-style shading or produce colors that feel less natural.

Full-color generation is possible, but it is not yet as consistent as monochrome generation.

I plan to improve this in a future version by strengthening and balancing the color training data.

After so many failed attempts and so much trial and error, this is the first LoRA model I have trained that I personally feel satisfied with. For me, this model represents my first real step into LoRA model creation.

Once again, I would like to sincerely thank 00X09901 for taking the time to help me despite all of my questions and the effort involved.