<h2>Foreword</h2>Hello everyone. This is the result of a couple of weeks spent on creating and comparing the same batches of images. I handpicked the models that produced the most desirable (or impressive) results, then I merged them. And merged them again. And again. The final model did that batch quite good so I'm satisfied. Some compositions resemble those made in original models, some were almost identical through all models I checked. What I meant is if you use my prompts with the listed models, your results can be very close. But they're never the same, so you'll find something new here even if you're familiar with all these models. And I bring you not one, not two, but three models. They have the same (almost, in case of v1.0) basic composition but with different weights. They are quite different, so read the descriptions. All models are given in two formats (pickle and safe) and two options (full and fp16 pruned). Pictures are done with the same settings across all three models. Sometimes, it leads to abominations, but this is not indicative of the model itself.<h2>Recommended Settings</h2><ul><li>Sampler: DPM++ 2M Karras;</li><li>Steps: 20-32;</li><li>CFG Scale: 10-15;</li><li>Clip Skip: 2;</li><li>Your favorite VAE. I like <a target="_blank" rel="ugc" href="https://huggingface.co/andite/anything-v4.0">Anything4.0's</a> the best, but it's too large. Another common option is <a target="_blank" rel="ugc" href="https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/vae-ft-mse-840000-ema-pruned.ckpt">vae-ft-mse-840000-ema-pruned.ckpt</a>. I'm uploading it here for your convenience. All the pictures in the gallery were made with the latter. Don't forget to rename it according to the model name (name.vae.ckpt).</li></ul><h2>Overview</h2>For this merge, my main points were beautiful eyes, controlled lewdness, variety of topics, attentiveness to prompts, compatibility with loras/embeddings, and, hopefully, colored eyebrows.<ul><li>Eyes turned out perfectly, and, when asking for glowing ones, this glow can become quite intense. It looks very cool and I advise you to test it for yourself.</li><li>Lewdness seemed to be perfectly controlled for me. The model is fully universal: it's equally good for sfw/nsfw stuff, and it generally won't show you boobs unless asked for. It can generate upskirts a bit more often than desired, though (and no, this behavior came not with the most obvious hentai models, color me surprised). Bonus model is generally lewder despite being more obedient in other matters.</li><li>Topics I've checked are cyberpunk/cyborgs, high fantasy, contemporary, abstract landscapes, skies, sea, and some lewd stuff, including tentacles. Almost all source models were not so good at least in some of these venues, but the result is good across the board. This time, I mainly focused on cyberpunk and cyborgs, hence the name, but, rest assured, it can do other things just as good.</li><li>Attentivess to prompts: this idea was born from my attempt at culling down my bloated prompts (especially negatives section). I use a group of negative embeddings as a last defense line, but other than that, the only thing that is needed there is the popular (worst quality, low quality:1.4) combo. Hands and fingers can be awful but there's not much you can do about it anyways. Everything else is at least as good as it was with my bloated list, or even better. In positives, I try to keep below 150 tokens, and it mostly goes well.</li><li>Loras/embeddings: for those I checked, it's quite compatible. Those that don't modify appearance work as intended, breaking nothing in the process. Those that do modify appearance are subject to usual limitations. This is not a photo realistic model by any means, so don't expect it to fully reflect your favorite actress, for instance.</li><li>Eyebrows mostly share the same color with hair. Not always, sadly, especially given that Stable Diffusion doesn't process tags like mismatched eyelashes or mismatched eyebrows. This makes me a sad panda, but it's a niche thing. And sometimes that can be inpainted or doctored. In 2.0 and Snakes the eyebrows are even cooler.</li></ul>Now, for obvious shortcomings.<ul><li>Leaking. Some colors or substances, especially when weighted enough, can leak from their intended parts and color something else. The most susceptible are nipples, areola, and privates. Don't ask me why. Judging by my experiments, this behavior was introduced by siiNCeys and Kuro Elysium. Their weight in the final model is not very big so I don't think this will be a very common sight. The only thing I can advise here is putting said color or substance in negatives. When asked to do directly, it still will use them, but it'll have less incentive of doing so anywhere else. Don't use tags like colored nipples. Despite being in the HD22's list, it does absolutely nothing. This problem is more prominent in v1.0.</li><li>Excessive detailing. Common to every Stable Diffusion model I tested. Not only those used here. When asking any body part to be detailed, you risk it to be detailed in a completely different direction. As in adorn it with gems or jewels. Yes, even privates. So don't use it outside of general tags like extremely detailed, intricate details and so on. Unless you are a fan of finding cut gemstones deep inside a pussy. Hey, I won't judge, you do you.</li><li>Fingers. This won't come as newsflash, but they can be horrible. Nothing much you can do, except hiding them out of frame, masking with gloves, or upping steps while praying to Tzeentch. However, he doesn't care about prayers, so won't be of much help.</li></ul><h2>Credits Section</h2>The models I used are these:<ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2457/siinceysmix-v1-and-v2-and-alternative-version">siiNCeysMiX Special Version</a></li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/5301/elysium-kuro-anime">Elysium Kuro Anime</a></li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/13135">MagicalMix v2.0</a></li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/12606/meinahentai">Meina Hentai</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/eimiss/EimisAnimeDiffusion_1.0v/blob/main/README.md">Eimis</a></li><li><a target="_blank" rel="ugc" href="https://github.com/Delcos/Hentai-Diffusion">HD22</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/scriche/Lamia/blob/main/bunyip.ckpt">bunyip</a></li><li><a rel="ugc" href="https://huggingface.co/a1n2n/Animelike_2D">Animeike_2D</a></li></ul>All these models are magnificent and I thank their creators. Some of them disabled any commercial use so it would be extremely impolite to not do this for my model. Regardless of the percentage of their work in the final mix, it isn't zero. If you need my model for commercial purposes, please don't contact me for permission. Frankly, I don't care how you use it. Check the list above and talk with those who've set these commercial restrictions. If you convince them, you're good to go. I require no credit whatsoever, but, should you want to mix this model, please honor the wishes of these people and do restrict commercial use accordingly. This is especially vital as in the short time Animelike_2D was here on civitai, its author explicitly prohibited any profit gains. So I can only reiterate this prohibition. Also, credit is now required. Not for me, I don't care about it, but to keep notes so permissions won't accidentally get lost.And, finally, there is a little issue with reproducibility of my prompts. I have --no-half-vae and --xformers in my launch arguments. The first one you could enable yourself and get the same results, but the last one is <a target="_blank" rel="ugc" href="https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/2705#discussioncomment-4024378">notorious</a> for bringing a little chaos in the system. Even the completely identical prompts may differ from each other. I've only seen it to affect colors of small areas but who knows.

Digital Soul

This is the final version of this model. I got everything I ever wanted from it, and any subsequent alterations will severely affect the looks. So I have to pull the lever here. Most highlights of this version include cleaner looks and more obedience when it comes to colors. Alas, it comes with a cost of robotic parts requiring more coersion. But in my eyes this is worth it. It still knows about them, so they never go away and may be weighted. Getting crisper looks without hires fix is way harder than that.

Files with SHA256: 8420b826ad5904982a775a715bfe3a89fc2371a031ee5c8e1d23146223579778