For V4
Like V1 it works with negative weight. It gives a bit different images compared to V1. I would not say it is upgrade. Both give nice outputs, it is just different. Weight depends on how long your prompt is and how many loras you are using.
You can use it without tags, but need higher weight. Usable tags:
traditional art, lineart, white background, simple background
Personally i like traditional art + simple background combo. traditional art alone works pretty well too.
You can use tags like monochrome to make it black and white.
There seem to be quite big variation between seeds.
For V1
FOR TRADITIONAL LOOK USE NEGATIVE WEIGHT. Weight between -1.1 and 0.7 seem to work best. Play around, prompt can alter the strength needed.
FOR MORE SOFTER STYLE USE POSITIVE WEIGHT
simple background, lineart, traditional media can make the effect more stronger.
This has been trained to work on pony models, can't guarantee it to work on any other XL model. Also results may
Description
Finetuning to remove some annoyance i didn't like
FAQ
Comments (2)
Holy, these loras are amazing! I would expect this style to have a serious versatility demerit, but if anything it seems to improve it somehow. Combining this with other loras and prompts that usually struggle to output coherent results makes them way more consistent and clear. And it always looks fantastic.
Style loras for pony in general help with consistency, but this one's a league above the rest. Obviously the style does have a lack of background and such, but maybe that's part of why it improves the model's ability; putting all the attention on the foreground.
Thank you!
This is not normally trained lora with dataset. It is a LECO. The training "finetunes" the knowledge of the model, so it is generally more versatile and compatible with other models/loras. It does not inject anything new to to the mess so to say, but enhances the models own knowledge :)
It does reduce the background, which is intended for the clarity. It does enhance the tokens simple background and white background, could try to drop those into negative to see if it helps.
Also can be used without the activation tokens, and from my testing seemed to give more clarity without losing the background. Might need higher weight.
Thank you for the comment! (～￣▽￣)～
Ps. Currently doing testing with background blurring to see if i can get the foreground more detailed with that. Drop a follow if you want to see more of these interesting experiments (～￣▽￣)～
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.