Charu Style

Charu style LoRA

Please read instructions as this is a fiddly temperamental LoRA

Sample images have workflows to use as a starting point, read on for further instructions on how to use. The sample images are all direct T2I generations using the LoRA, upscaling and inpainting can significantly improve the quality.

Experimental LoRA originally intended only for coloring in real art, still needs some work to get good results out of it.

How to use:
LoRA strength
1, Never has to change from 1 as the trigger word actually works.

Trigger word
charu_style
Strength 1.0-1.6
Trigger word is required, without it the LoRA does little to nothing to the image. Start at 1.0, adjust this up last.

Genertic positive prompt
masterpiece,best quality,amazing quality
Strength 0.4-0.8
This improves the image quality, but makes the built in style override the LoRA, so the strength must be turned down depending on the image. Start at 0.8, adjust this down first.

Generic negative prompt
bad quality,worst quality,worst detail,sketch,censor
Strength 0.6-1.0
As with positive prompt, required for good image quality, but has to be dialled down or it can overwhelm the LoRA. Start at 1.0 and adjust down if adjusting the positive prompt strength isn't enough.

Play with the strengths of the trigger word, generic positive and generic negative, the required values will vary depending on the image. I generally start with all at 1, lower the positive and negative strength first, turn up the trigger strength if it really won't play ball. Watch the hair highlights, his lightning style highlights are the first thing to go when the LoRA struggles.

Background / Why is this LoRA so fiddly:
The source dataset was every original color image I could find. These images are few in number, all have similar composition (1girl, standing, naked, blush, outdoors, etc), and many are of questionable quality. I removed the worst quality images, decensored them all using inpainting, and blacked out the backgrounds (many of the backgrounds were postprocessed photos or similar which destroyed the model). Backgrounds are blacked out rather than alpha masked because alpha masking that much of the image pushed the model to generate extra limbs and such. This got it as far as generating decent girls, with backgrounds that didn't look like they were melting.

Because the dataset images are all so similar it naturally learns the composition and content as well as the style, meaning that the further the generated image was from the training set, the more it would lose the style. Basically, if it wasn't generating a naked girl standing outdoors looking at the viewer, the style vanished. To fix this it is trained with a regularisation dataset. This makes the trigger word actually work and be required, unlike most LoRAs, allowing for easy control of the intensity of the style just using trigger word strength.

The final issue is the generated image quality. Without generic positives and negatives the image quality is generally rubbish, but with them the model's built in style overrides the LoRA, unless you turn the LoRA style up so strong that the image melts. The solution is to apply a lower strength to the entire generic positive and negative prompts. For every image there is a sweet spot where they are strong enough to make a good image, but weak enough that they let the LoRA style through. It tends to have strong inflection points, where the positive prompt might overwhelm the LoRA completely at 0.8, but it works fine at 0.75.

Long term fix would be to use this LoRA to color original lineart to create a synthetic dataset with more image diversity.