Pale Mix

A merge of DonutHoleMix and BE-Mix (removed from Civit, no idea why) along with Gothic Punk Girl and Hitoimim art style (artist link) loras baked in

I found that using the mentioned loras with specific strength on BE-Mix resulted in an interesting look, this is that just merged into a single model.

DonutHoleMix was more of an afterthought, and really only because I had to use a second checkpoint to do it in ComfyUI which I use exclusively, but I found that its CLIP model makes the results better and it allowed for a more anime-like model as well

Gothic Punk Girl Lora (strength: 1) helps make more detailed clothing and makes characters cuter, no matter whether you're going for an explicitly gothic look or not

Hitoimim art style (v1.0, strength: 0.6) adds more detail to everything, eyes, hair, clothing, patterns, lace, frills etc. Also dramatically improves some animal body parts like tails or ears, as well as animals themselves. Can sometimes add animals unprompted.

A side-effect of both loras is making the skin more pale than normal, hence the name.

There is also Hipoly 3d Model LoRA (v2), but only at strength: 0.1, so its effect is negligible

The Realistic model should be considered to be the main one. The Anime model doesn't carry too much weight from the loras which are essential to the effect in the realistic model. It is different enough from the realistic model or any of the component models, however.

The Realistic/Anime distinction isn't very accurate, both are more on the 'realistic' side (though maybe it should be 2.5d/3d instead?), the anime one is just 'more' anime and vice versa

Both checkpoints work with booru-style prompts.

Realistic Version:

input: 100% DonutHoleMix

middle: 100% BE-Mix

output: 100% BE-Mix

Anime version:

input: 100% DonutHoleMix

middle: 100% DonutHoleMix

output: 50/50% DonutHoleMix/BE-Mix

The Loras are only mixed into BE-Mix before merging. DonutHole's U-Net and CLIP are unmodified

The CLIP model is the one from DonutHoleMix (modified with lora)

The whole process is done in ComfyUI using the block-merging node

For the exact recipes, load the checkpoint into ComfyUI via drag&drop or load it as a ComfyUI workflow

VAE: kl-f8-anime2 or vae-ft-mse-840000-ema-pruned for both models

Tips:

• Since there is much focus on fine detail, the results may not look too well on lower resolutions.

The showcased images were first generated in resolutions such as 576x832, 640x960, 640x1024, 512x1280, 896x576, 1024x640, etc., then upscaled 2 times into a second pass with 0.35-0.5 denoising (or hires-fix for AUTOMATIC1111 webui users) followed up with face restoration, then upscaled 2 times again.

• Don't go too high with denoising on hires-fix. It messes with face structure, especially at an angle, making them uglier. Even 0.3 can visibly change eye/mouth/nose shape. In my taste, always for worse.

ps: don't mind my prompts in example images, It helps with organising them to get the framing I want, I swear