Quasarcake版本v1.0 (ID: 104212)

Style LoRA trained on Quasarcake's style. Should work on most models with an anime bias, but the effect will be strongest on Anylora. I recommend pairing it with epi_noiseoffset LoRA to strengthen the contrast, and blessed vae for the same reason.

Recommended settings are as follows. If you're on Automatic WebUI, you can paste it in and save it as a preset style so you don't have to navigate as many tags.

Positive prompt:

mksks style, {prompt}, (masterpiece, best quality, high quality), <lora:Quasarcake768:0.8>, <lora:epi_noiseoffset2:1.5> 

Negative prompt:

(worst quality, low quality, normal quality), (colored sclera:1.3), {prompt} 

Generation parameters:

• LoRA Weight: You can technically go up to 1.0 weight, but 0.8 is where you get the best balance imo.

• Config: 7.5

• Clip Skip: 2

• Sampler: Any, but all of the examples are DPM++ 2M Karras

• Width/Height: Since the model is trained on a 768x768 dataset, it can handle higher resolutions pretty well, and lower resolutions tend to come out more "zoomed" than normal because of it (sometimes works out in your favor). All the examples except one are generated at 512x640.

• Highres fix?: It's fairly stable, so you can get away with using latent if you like most of the time. Either Latent upscaler <=2x and <=0.6 denoising if it's single subject, or R-ESRGAN 4x with <=0.75.

• Adetailer: The original style is kinda sketch-like around the eyes, which SD does not like. In the end, that results in blurry eyelashes. I can recommend running with Adetailer, if you find that to be an issue. This is especially useful if you're generating at lower resolutions due to VRAM restrictions.

描述:

训练词语: mksks style