Deep Dark Hentai Mix - NSFW Anime

https://civitai.com/models/221751?modelVersionId=256648

Best settings: Sampler: Some sampling methods don't work with this model. I typically use any of the following settings, all of which work great:

Euler a DPM++ 2S a Karras DPM++ 3M SDE Exponential DPM++ 2M SDE Karras

Steps: Anywhere between 30 and 50, depending on which version you are using and if you are going to run hires fix or a refiner.

CGF: Give it a low CFG so it has the freedom to create some really interesting things. I typically use between 6 and 9, but I have gone as low as 3 and as high as 12 or 13 before images start looking burned.

Prompting: If you want to improve the quality and resolution of images, use tags like high resolution, highres, best quality, 4k. Intricately detailed and amazing details work to add extra complexity. Cinematic pushes images to 3D and makes some generations realistic. Don't use masterpiece, immaculate, absurdres, absurd res, 85mm -- they all do more harm than good. No negatives are needed as a standard prompt; things like low quality or bad anatomy aren't necessary. But if you are aiming to generate NSFW imagery, adding "children, loli, child, kids, teens" to the negative is a good idea -- just in case.

Version 4.0 I had been working on version 3 of Deep Dark Hentai Mix for quite some time. A few days ago, I was still not quite happy with it, so I was still making small tweaks and adjustments. But while working on Suzanne’s XL Mix version 6, I figured out a different way to pre-train some of the concepts into the models before I merged them. I tried that here and discovered it worked for anime as well! Much like Suzanne’s mix, I merged it with what was going to be the third version of Deep Dark Hentai Mix, and I’m really happy with the results. So rather than tweaking version 3, I'm just going to go ahead and publish version 4. Deep Dark Hentai Mix version 4 has the same composition and much of the aesthetic of earlier versions but a much greater range of understanding of concepts (both NSFW and SFW), as well as fewer pesky watermarks. However, prompting is different than in previous versions, particularly in generating NSFW images, as it responds more to descriptions than previously trained keywords.

Version 1.2 Version 1.2 uses the most up-to-date models and has a few different tweaks to get improved images, follows prompts better, and understands a few more concepts than previous versions.

Version 1.1 Version 1.1 is the latest version of the earlier, more anime-focused Version 1. It is made of the same models, but the mixing method is slightly altered, as are the model weights. If you like the style of Version 2 over Version 1, this isn't going to be your cup of tea, but don't worry; there'll be a new version improving on Version 2 soon (part of the method refinement that happened here was necessary before I could move on).

Version 1.1 does some things better, and a few things are not quite as well as the earlier version. Prompting (in my opinion) is improved, but it takes an extra generation or two for me to get images with exactly what I want. Lower CFG values seem to produce better images, so sometimes, it's better to start lower with a simple prompt to see what the model comes up with. It does very well with NSFW imagery with very short and simple prompts.

Version 2 Version 2 was mixed using a new custom base model and with a slightly different formula. I'm not sure it's better than Version 1, but it definitely does some things better -- so I guess it's improved depending on what kind of images you are going to use it to produce.

Generally speaking, v2 seems to understand prompts better, produces fewer horrors from beyond, and has less random, unprompted hardcore sex. The artistic style sits between anime and semi-realist, and you can usually prompt it to go either way. It seems to understand prompts better for more SFW art, but as a balance, it needs a bit more coaxing to go to extremes in the other direction. It also likes to produce really wet images. Lots of fluids on all kinds of things. Not sure why.

Version 1 I hadn’t been able to find an SDXL anime model that had the exact style and aesthetic of some of my favorite 1.5 models. So I set out to do a combination of merging and training to achieve something similar for SDXL. The results of the first attempts produced only images of horrors from the abyss, but after a few tries, I achieved something that was semi-stable, produced nice quality images with some well-thought-out prompts, and had the look I wanted. The only problem was that it was also extremely horny. It removed clothes from subjects that didn’t even have clothes. It somehow interpreted every prompt in the most sexual way possible (and in some cases in ways that were not possible). While trying to fix the issue, I accidentally merged it with some of the earlier failed model attempts and found that, oddly, the mistake both somewhat cooled things down (so it mostly only produces NSFW content on request) and produced a mostly stable model. Thus, the Deep Dark Hentai Mix was born.

The Deep Dark Hentai Mix model will mostly produce anime-style images when prompted with Danbooru tags (1girl, 1boy, etc.) but will produce fantasy art and semi-realistic styles for some prompts in natural language. It can do SFW images but really doesn’t like to make them. It really prefers NSFW content and will happily produce a wide range of softcore and hardcore images. Without careful prompting, pants will be missing, anatomy will be on display, things could be inserted, etc. You’ve been warned. Forcing it to generate images on prompts it doesn't understand may result in images that will haunt your nightmares.

Credit to the following creators for some of the models used in this merge: diamond.visual.XL, ShikiAnimeXL, SereneXL, Hentai Mix XL - Road to freedom, DeepBlue XL, Pyro's NSFW SDXL. I’d need a degree in calculus to produce the formula used, but it doesn’t matter because trying to recreate it in three different merge programs in Automatic1111 and ComfyUI all made slightly different models. But it involved training, adding model differences, and merging by block weight.

by Dark_infinity