Note: The latest is not always better.
I aimed to create a merged model that coexists with the anime drawing style while compensating for the background art writing that is often lacking in illustrious derivative models.
This model is characterized by the fact that it does not collapse on reasonably small faces without adetailer, so it recognizes faces even in small areas when adeteiler is used.
If a long prompt with a large number of tokens makes your image look grainy, like it has overly sharpened noise, try lowering the CFG value to around 5, or lowering the CFG during hires fix to around 4 to address it.
Additionally, switching the sampling method to something like Kohaku_LoNyu_Yog can help mitigate issues that occur with high CFG values.
For a more fundamental fix to overly sharpened noise caused by long prompts or high CFG, consider installing the “Dynamic Thresholding (CFG Scale Fix)” extension. Try enabling it with Simple Mode, setting Mimic Scale to 7 and Threshold Percentile to 1.
Recently used value: CFG scale=around 4~6~8~10
Sampling method=Kohaku_LoNyu_Yog
Schedule type=Normal
Sampling steps=about 30
When a thick coat of paint like an oil painting style is required.
Recently used value: CFG scale=around 4~7
Sampling method=DPM++ 2M
Schedule type=Karras
Sampling steps=about 25
Width:Height = 1152x864 864x1152 1280x960 960x1280 1024x1024
hires fix:
Upscaler: ScuNET PSNR
Denoising strength: 0.15~0.35
Upscale by: 1.389(1152) ~ 1.25(1280)
Hires CFG Scale: around 4.0
The prompts I use most often are at the end of a sentence.
very awa, masterpiece, best quality, newest, highres, absurdres, very aesthetic, (dramatic lighting, cinematic lighting, perfect lighting:0.5),
negative prompt:
(bad:1.6), error, fewer, extra, missing, worst quality, low quality, (bad hand, bad feet:1.2), (6fingers:1.8), (fewer digits, 6toes, 4toes:1.7), (bad anatomy, conjoined:1.8), (wide hips:1.8), censored, simple background,
The output image may have a large hips. If this bot