**I'm a beginner and I've just made LORA for the first time. There may be a lot of unsatisfactory parts.

**I tested it with several 'LORA', and a young woman and a slim chin 2D photo appear as a result. There was no data on men, and when learning data, there seems to be a problem because of the lack of a woman's face. See this

My Lora tag is 'Chinese Sketch 002', but the numerical values ​​can be directly reflected with a gauge bar.

**Here are my recommended settings:

I use the 'NTRMIX XIII' model.

'Chinese Armor?'=0.8, CFG=5, STEPS=25 or 28,

Sampler is recommended as 'DPM++ 2M karras',

and if you want to reflect additional roughness,

use another Lora called 'ScribbleRay (SDXL)'.

and, If you want a dirty look, try adding '薄塗り / USNR STYLE' here. (I don't know if I'm the only one who gets this effect)

(CFG=3.5 for rougher lines)

- Some test results <2025 02 26>

**If you want rougher lines in my Lora, you can use another Lora, 'ScribbleRay (SDXL)', here.

Sampler recommends ‘DPM++ 2M karras’.

CFG=3.5 or 5 or 7(3.5 or 5 recommends), Step recommends 25(CFG=3.5 or5) or 28(CFG=7).

The value of Chinese armor? is also possible with 1, but ‘0.7~0.8’ seems to be safe (testing recommended from 0.8)

- Test when the value of Chinese armor? is 1

When the value is 1, the probability of the face being damaged is high at CFG=7. It doesn’t matter at CFG=3.5

At CFG=3.5, the Lora drawing style I want is almost implemented, but the intensity of the prompt reflection is low

At CFG=5, Lora's drawing style and armor features are reflected appropriately. However, the prompt reflection is not always consistent.

At CFG=7, Lora's drawing style is weak, but the armor features are implemented moderately in 2~3 out of 4 pictures

Based on the Chinese armor?=1 value, experimental results,

If Cfg is below 5 and Step is below 28, there is a high probability that the fingers and the picture will be distorted. I don't know the appropriate level.

Therefore, it is better to exceed Step=28.

CFG is the intensity of prompt implementation,

- CFG=3.5, Lora's intended drawing style is almost the same

If you want rougher lines in your Lora, you can use another Lora, 'ScribbleRay (SDXL)', here.

(The effect is visible at CFG=3 or 5)

However, it seems that having a large number of prompts or dynamic prompt expressions is disadvantageous because the picture is likely to break. (Maybe because the CFG value is low?)

- CFG=5, much cleaner lines, but Lora's slightly light colors and very fine line expression

- CFG=7, the lines feel cartoonishly clean and thicker. It feels like only the clothes and armor designs are appropriately reflected

Conclusion)

Lora's rough drawing style is hard to reflect in 3.5 or higher, but the armor features are implemented to some extent in 7, so

If you need a rough drawing style, CFG value is 3.5, and if you only need the armor and clothes features to some extent, 5 or 7.

*However, the lower the CFG, the lower the prompt implementation intensity, so you have to think carefully

(There is a high possibility that the design requested in the prompt will not come out)

*Please let me know if you have any new opinions.

This is just my practice so please be careful when applying it. I wanted to change the drawing style, but I'm not good at it, so I'm not sure if the first LoRA I made will work properly. If the face isn't displaying correctly, including the bibliography author or font in the prompt may cause the face to display correctly. Ultimately, the drawing style seems to vary depending on the LoRA, model, and prompt.

The effect I want is to change the drawing style like the picture posted. So please use good judgment before applying it.

ps. I tested LoRA using the 'Illustrious-XL' OR 'NTRMIX XII' model and it seems to work as expected to some extent.
ps2. When I applied my LoRA with 'Stable Diffusion's AUTOMATIC111', the image was not selected as well as I wanted, but strangely, it seemed to work when I applied it in 'Civitai'. I'm not sure why, maybe because I'm a beginner.