interesting part is that this picture is pretty because i set up stable diffusion wrong.
desaturation was partially caused by "auto" in VAE settings. After I set it to nai.vae, it became more saturated (still good tho: https://files.catbox.moe/850cnu.png ). You can try disabling VAE (set to "auto" or "none" in settings) for more desaturation.
More negprompts to the god of negprompts! Also I didn't know you can just describe the image, I thought you need to just list tags and hope model will understand what you want.
@antlers_anon are you sure there's no hypernetwork you've selected in your settings tab, or TI that's being used to generate this image?
I tried to generate this with your provided mix__2.ckpt with all of the exact same settings (prompt, sampler, CFG scale, seed) and it still results in a fairly realistic non-anime style.
If nothing else, would you be willing to share your embeddings folder and hypernetworks folder? (just a screenshot would probably do as well?)
edit: looks like the firstpass width/height was set to 0x0, so it was just straight-up going for 768x1280 for initial resolution. I ended up finding that it gets pretty close, especially if I use kl-f8-anime2.vae instead of novelai/anythingv3 vae.
I'm guessing that colour correction and some filtering adds a little noise
The model used was a 50% merge of Protogen x3.4 and Anything-3.0. Believe it or not, but the "aroused" and "horny" keywords are there for the facial expression. "Dreamlikeart" was included by mistake when I used an old prompt with a different model (the previous model I used was a 50% merge of Dreamlike and Anything-v3).
