https://civitai.com/models/398855
I've tried to merge an anime model that will look as detailed as possible, yet remain vivid and clean, with no realistic features (nose, nostrils) on character's face. Also I've tried to make character's skin realistic'ish: soft and smooth. Prompt versatility is also a must for me. So for now I came up with this merge. That's my first time using SuperMerger, I still don't know a lot of things about what block does what, but I did my best at the moment. I hope you'll like my mix. That's not a final version though, I'll work on updating it in the future.
All preview Images made with Hires and ADetailer.
Recommendations:
Use danbooru tagging for prompts, THIS extension will help you with that;
I'm using kl-f8-anime2 VAE, but if you find it too colorful, feel free to use any of your favourite ones;
Set Clip Skip to 2, it's very important for anime models to work properly;
Samplers: Euler A, 30 steps, for fast, stable and cleanest generations; DPM++ SDE, 24 steps for more details, but generation time will increase significantly;
Upscaler: 4x_fatal_Anime_500000_G, is the best IMO, Upscale by 2, 15 steps for Euler A, 10 steps for any other sampler, 0.3-0.4 denoising strength;
CFG Scale: 6-8, my preferred scale is 6.5;
Try ADetailer extension for drastically better faces;
For prompt base I'm using (best quality, absurdres:1.2) in positive, and (worst quality, low quality:1.2), (blurry, bokeh, dof, fog:1.2) in negative;
Use Async's Negative, or EasyNegative embedding, I prefer to put it on 0.8 weight. The first one will give you more sharp, contrast image with slightly better hands. The second one is softer, with more volumetric lighting. Remember not using it with Style LoRAs, because it may ruin the style;
You can use Bad Hands 3 for better hands, just don't rely on it too much;
Also try realistic, 3d tags, those are really good. Be careful with weight though;
Hairdetailer will help you make amazing hair. It'll also help to turn contrast down a little bit. Just remember to adjust weight to around 0.6, because this lora affects model's style.
Credits:
Huge thanks to TroubleDarkness for teaching me everything I know about Stable Diffusion. Go check out his DivineElegance, it's amazing;
Thanks to Kybalico for making StrawberryShortcake;
Thanks to technobyte for making Milky Wonderland;
Thanks to You, for trying my mix. Have fun!
