>>11282 Ah also, I used a Succubus Covenant [Pony] lora for the above images. https://civitai.com/models/432436/succubus-covenant-or-game-cg-style If you're curious about the specific generation parameters, they're embedded in the metadata of the images.

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Optimum-SDXL-Usage Didn't know this existed but it is also very useful.

>"being employed to do office work by the elf OL boss in another world" these isekai titles have been getting awfully specific huh?

>>11285 though not as good as "conquering heaven with my bitchy demon girlfriend"

>>11286 I like how her bright skin masks the usual AI glow/oversaturation. Aside from looking very carefully at details, I'd have problems judging if it's AI

>>11282 Can AI do extremely specific fetish stuff in the style of a particular artist yet?

>>11290 Depends heavily on what you mean. There's a chance that one of the models understands that concept, or you might be able to find a lora for it or related concepts on civit. Alternatively, you can even train your own LORA. I was able to train one with only a few GB of VRAM but SDXL loras might be a bit harder since the models are much larger. I'll look for a good SDXL lora guide in a bit.

>>11292 Also, you can mix and match loras with varied results, but it seems better lately on SDXL/Pony. So if you can find or make an artist style lora, and then use a fetish lora, you should be fine.

Not exactly monstergirl though she is a humonculus. But here are some results by messing with Asanagi, then Hews style lora.

>>11286 I'd prefer "conquering hell with loyal angel gf". So Helltaker with some extra elements. >>11297 That's totally Kuro and Illya. But it looks pretty good, the technology came a long way. Pretty soon commissioning a specific artist with a notable online presence wouldn't even be necessary, as you could just train a model on his pics and have unlimited free generation. I wonder how the necessary amount of training data will change in the future. As in, how few pics would future models need to understand a specific artstyle or how to draw a specific object. Right now human artists need only a few pics of a specific thing to know how to draw it, and they can probably make do with a single pic when it comes to artstyle (tho of course more would be better).

>>11359 yeah, of course. Both Illya and Kuro are homunculus, i do enjoy fucking with around. Even generated a few Hataraki Ari based Chloe pics, may upload it later

>>11364 do post them while at it, also tell what artist are imitated in pics 1-2 and 3.

I genned some using NAI Anime v3, but they might be a little heavier than some anons prefer but hopefully someone likes them.

>>11367 and two more, I have a lot but these six are the best I feel.

>>11366 I just mixed Chloe with Felicia from darkstalkers. Didn't specify what artist it should copy

>>11282 >SDXL model called Pony Diffusion. Is there a quickstart guide to what I need to start doing this?

>>11404 I basically tried to make the thread introduction as close to a quickstart guide as I could. I might write a proper one when I have time. That first github link explains pretty well how to install A1111, but if you have any problems feel free to ask and I'll try to answer when I check the thread. A1111 has been having a few problems with SDXL for a while though, so some problems are expected. There are like 2k "issues" / bug reports on the github so I'm pretty sure the dev just gave up trying to look at them all. Once you've installed A1111 webui, you need to install a version of CUDA (if you have an NVIDIA GPU, no idea about AMD) so that your GPU can do AI calculations. https://developer.nvidia.com/cuda-12-1-1-download-archive https://developer.nvidia.com/cuda-11-8-0-download-archive 12.1 should be the correct version, but other installs have used 11.8. I think SD A1111 by default today uses Pytorch with CUDA 12.1 but I'm using 11.8 at the moment because it's more stable for me. Try 12.8 first and if that doesn't work ask here because that's a hassle. After that you literally just download a single file (the pony model or another one) from Civit and put it in models/stable-diffusion and start up the webui, then load it in through a model switching tab in the top left. Go to the txt2img tab and prompt anything then it should work. If it gives you an error check the console and it's usually a vram or memory issue which means you need to use the optimization args linked in the thread or add/increase a system pagefile (harddrive memory). There's another alternative that is supposedly more stable and has higher performance, but I don't think it has any or as many extensions and has flowcharts instead of GUI. Didn't mention it at first because it's probably only better once you have a good understanding of stable diffusion. https://github.com/comfyanonymous/ComfyUI

>>11427 12.1*

>>11427 I will get around to this soon. At least buying nvidia finally paid off.

It's everything anyone could have wanted in a monster girl.

It's remarkable how often I can start something only for the site to go down into maintenance.

>>11282 >It automatically detects faces in your picture and inpaints them to keep them from looking terrible. Any advice on using this?

>>11454 Not sure what you mean by "using this" so I'll just cover everything I think of. First, you install it in the extensions tab using the url of the github repo and then restart your webui. I recommend just installing all the extensions before restarting or you have to do that multiple times. In txt2img where you generate your images, you'll find the adetailer tab below the seed value. You are given four tabs to work with, and on each you can choose a "detector" model. The detector model will look for specific patterns in an image, crop them, inpaint/fix them, and paste them back onto your output. SDXL usually only needs help with faces. I have found that the default model, face_yolov8n.pt, works well enough for this. Below the detector model is the prompt for inpainting. It works just like the normal prompt. Usually the default works fine but it's good to use a specific prompt if you want to highlight a specific detail during inpainting. There are face, hand, whole body, and I think a few other types of model depending on what you want to inpaint. The other tabs don't really matter but I'll go over them briefly. Detection - this determines what the model considers to be a part of what you want to be inpainted. The threshold determines how much like a "face" (or whatever you want to find) has to be for it to be cropped. The ratios and offsets allow you to change what is cropped by the ADetailer if you're looking for something specific. Inpaint- You can try messing with this if you get bad results. You can change the denoising value (essentially how much your original output should be overwritten when inpainted, higher is more) and also resolution to see if it changes your results in a way you like. ControlNet - if you're doing a full body inpaint, you can try to use controlnet to keep the same positions after inpainting, otherwise it's much less reliable. It basically feeds the original image as a reference to control net and then inpaints with that. If you're interested, see what I mentioned in the top of the thread about controlnet.

>>11457 https://github.com/Panchovix/stable-diffusion-webui-reForge Also, I ended up switching to this today. Not sure if I recommend switching to it for anyone else yet but at the very least it is far, far more stable than A1111 for me. A1111 crashes constantly on my PC, I originally planned to mention webui-forge, which was the original version of the above, but it has apparently been discontinued from its original purpose recently and will now have a lot of experimental features (and therefore bugs), so I never tried it. Reforge has no crashes (yet), I can actually use my pc while sd webui is generating 4 images at a time in the background. A1111 used 50GB of ram due to a memleak that hasn't been patched for some computers since SDXL support was added. Only downside so far is that some extensions seem to be broken, fortunately civitai helper and adetailer still work.

Made an experimental lora for black alice. Took about 3 hours to train. It's far from perfect but at least it gets the outfit and general appearance right. Faces don't go very well but with ADetailer you can smooth out the problems. https://gofile.io/d/AlAVef

>>11463 I have discovered that the earlier versions of the lora actually provide better faces and seem to be a little better. If anyone expresses interest I'll upload those some time later today.

>>11464 Wow, that's an amazing looking Black Alice, ignoring the fingers. Shame that I have an ancient PC that can't into AI.

>>11457 I mean I can see it doing the painting but I'm not even sure if it's changing anything.

>>11466 If the faces are already pretty good it probably won't appear to change much. If you have live preview on (showing you something while it generates) then it'll show you if it detects a face, drawing a square around it with text like "face 0.8" (80% certainty) and then it'll inpaint. If you're looking for something specific try changing the adetailer inpaint prompt, denoising strength, and resolution. If it just flat out doesn't detect the faces try changing the lowering the detection threshold or changing the detection model.

>>11474 Sometimes the face can also be corrupted by the LORA. The black alice lora has that issue above, probably because I over trained it. I made an Ilias one after uploading since they take like 20 minutes to build a dataset and 3 hours to train in the background. Not ready to post that one yet but it had the same issue. If you use a custom prompt it won't use the lora for the inpainting unless you specifically put in which may have better results You can also change the lora weight by changing the number in the lora identifier Ex: <lora:monstergirl:0.8> vs <lora:monstergirl:1> That can reduce deformations caused by the lora

>>11475 On further reflection the face issues on mine might actually be my fault. I'll probably try adding close up shots of the faces from the images to the training data. With only full body shots I guess it makes sense the faces would be bad. Still looks good enough right now with inpainting though. I'll release my current one or a new one I train tomorrow if it looks better.

>>11476 Sorry I've been very busy lately. I couldn't get it working quite right but I think it's even better with adetailer now, and the body. Using 2d and the "detailed face" tag probably makes it better. https://gofile.io/d/IJPyHx It's trained in epochs which you can consider a measure of training time. The ones without extra numbers are "less trained" which can reduce artifacts from over training. Each higher number is trained for longer and the one with no numbers are trained for longest. I included the last three because I couldn't decide which one I liked best.

All of those were made using Novel AI and their "NAI Diffusion Furry V3" (I know, I know, Furry bad yadi yada but the model is very sound) I am still very much a beginner yet I do enjoy the results after messing with it for less than 48 hours. (They are to my taste so it might not resonate with you, I like the angery and rapey kinds)

>>11676 Looks pretty good to me. I've never had too much interest in models that can't be run locally because it feels weird having corporations spy on your usage of the models I've noticed this thread has stagnated a bit, probably because of the high hardware barrier for entry and the uncanniness of AI sometimes. Hopefully, that'll change soon enough. A new AI image model released within the past week or two called Flux. On it's own, the model is too big to run even on a 4090, but it's structured in a way that let people make optimized versions with mimimal quality loss that can even feasibly run on a gtx 1060. Seems even better than Stable Diffusion by a noticeable margin, and in a completely different league with understanding prompts and accuracy. Only problem is it lacks finetuning for anime and has no nsfw capabilities, because it seems they either intentionally avoided it or tried to prevent the model from making it. But people have already gotten it to generate naked people, but not sex. Give it a few months and someone will probably make loras for it.

>>11688 Thank you! I was in the same boat too, until I realized that not everyone was rocking a 14900KF/4090 with 64 gigs of ram while also rocking some fierce autism and wizard powers. (let's be real, to have the real good results it's a rather arcane and in depth skillset) Since this is mostly used by fury, I'm quite positive that even if someone checked the stuff I make I would be the vanilla fag, it's still very annoying but the results are there, if you're into niche stuff it's probably your best bet before either Flux or Intel release Lunar Lake and Nvidia Blackwell, those are dekstop solutions with dedicated AI compute units on them.

>>11693 I'm not sure how good my stuff is, really. It's certainly much better than when people first tried doing nsfw on Stable Diffusion. I think I'm desensitized to small artifacts because of how bad it was for nsfw and anime when sd 1.0 came out. **Warning and disclaimer: attached picture may or may not irreversibly kill any and all boners. It's like if xelvy came back from the dead and his creations inbred for ten thousand years and fell in a vat of image artifacts. You have been warned.** And no, I didn't ask it to generate multiple heads. It kind of just used to do that. Also, updating on Flux, people have already created nsfw anatomy loras, blowjobs, and anal. Surprisingly no cowgirl or other vaginal sex yet. They're moving impressively quickly. I'll post a tutorial on how to install Flux (once I figure it out myself) when I feel that the NSFW makes it worth it to switch or dual run with sdxl. Blackwell chips are completely a joke, don't even bother with them. Modern NVIDIA AI chips are sold for over 10x production cost. Blackwell chips are estimated to price at around $30-40k msrp, and you have to get a quote from NVIDIA, can't just buy it because they're corporate chips. Your best bet is waiting for the 5090 release next year if you're interested enough. That's what I'm doing, but I'm afraid scalpers will bot them all before I get my hands on one. I'd say AI is a bit less autistic to install locally than before, the biggest problem is just that python is probably the most insufferable software I have ever used, bar nothing. I'll forever be sad most AI is built on it. Maybe I'll make a guide at some point on how to deal with python environments and bs. Half the errors I've had with AI are exclusively because python is awful.

>>11699 Me neither, I'm a starter with this model and this prompting method, I assume it's quite good since I was able to make that while barely trying, it's been 5 days and the results are quite good if I dare say so myself. As meymey as it sound, I always had troubles with offline models and for whatever reasons I tend to break anything coming from GitHub, for a rather cheap price I have unlimited generations with this service, already made 2000 pictures, none of them were trash (I can't share most of them here, my tastes aren't meant for this place) I would like it very much if you could somehow streamline the whole Flux process please, I'm as surprised as you are with the lack of cowgirl/riding. Sorry, that's what I meant, the 5090 and other consumer grade hardware, for many reasons I have a Quadro at home and it's... finicky with AI and in particular with Stable Diffusion. I agree with you, Python (mostly PyTorch really) account for half of my problems, for me there's a before and after GitBash methink, but still one day, perhaps we will get all the dependencies in a single .exe and not have to hunt down some arcane .bat from questionable website :^)

>>11702 Yeah, still looks pretty good to me. I just don't know if that's because I'm desensitized to ai or not... I don't blame you for having trouble with some of the Github installs. Sometimes I'll pull updates from a github repo and a commit will have irreversibly broken a feature and contributors won't fix it until weeks later with an "oopsie". A lot of times changes aren't actually tested so the userbase is used as quality assurance. The tard rage I feel when I have to spend hours fixing those errors myself makes me think I'm autistic. Good thing I know a bare minimum of python coding and pip, but I can't touch any of the AI related functions. Theoretically it's probably possible for me to pack stable diffusion webuis into exe files for everyone, but it would have to have a pytorch version packed into it built for a specific version of CUDA. That pytorch version in the exe would A. have to be compatible with your pc B. have a version of cuda compatible with the graphics cards of every user. I'd probably have to make an entire FAQ about how to tell if you have to manually setup the webui The exe would also likely be over 3 gb because of all the python dependencies built in (especially cuda baked pytorch) https://civitai.com/models/666326/first-edition-sex-positions-rough-versionnsfw-flux-lora?modelVersionId=745718 Someone finally released vaginal sex positions for Flux today. I think it's worth me at least taking a look at now. Flux only does realistic people, but one of the first loras someone released was an anime style lora so it doesn't matter. Not exactly sure how monster girls would go. Anyways, I'll take a look at it and maybe post my findings when I have free time, probably this weekend. Perhaps even an arcane bat from a questionable capybarachan thread will be coming to an area near you! Or not... depends on how hard I think it is to configure any of this, or if I can get it working. I do recommend the RTX 5090 when it comes out, but even the RTX 2080TI or 3080TI can work well these days. You can get the former for $200-400, and it can generate a batch of 8 Pony images in about 10 seconds, excluding adetailer inpaints. I assume the 3000 series will also rapidly depreciate in value after the release of the 5000 series.

>>11708 Unfortunately, I cannot get FLUX working yet. Works fine enough alone but occasionally has crashes that seem to stem from being out of memory. LORA usage spikes memory drastically, and a single LORA takes 50 seconds to load for me per generation, likely because of how Flux works differently from SDXL. It seems we need to wait for one or multiple things A. Community made Flux variant model with built in NSFW and better concept support so you need less loras B. More optimized, stable lora code that does not crash and works quicker C. $20 RTX 4090s Flux probably could work right now but unfortunately for me I only have 32gb of vram and a 2080TI. Pity. Hopefully it's more optimized soon, I'll check again in a week or two.

>>11709 32gb of RAM* I do recommend switching over to Forge as they have stabilized and finished the breaking changes they did previously. SDXL should work great on it. It is also the only webui that I know of that currently supports flux (unless you count ComfyUI but that's less user friendly), but forge's flux support is, well- as I've said previously. The install process was actually remarkably easy. I just download a zip and ran a batch file and it didn't even have to download everything because it comes with a localized python environment in the zip file. https://github.com/lllyasviel/stable-diffusion-webui-forge Pony SDXL should work easy on stuff with 5-6gb of vram.

>>11708 That's pretty neat! I would enjoy very much to see that, my only "skill" is being able to screw with RPGmaker, perhaps with a good enough model I could make "scenes" for a potential game, maybe even involved the anons on capybarachan? that would be cool. A 5090 with 64 gigs of ram is on my bucketlist once it's released, I don't know what CPU I'll go for, depends if the 9950X3D is better at chugging along or if the Bartell lake from Intel with 12p cores and a 8400mhz+ of ram will be better, we will see. >>11709 That's a bit sad, but it was a bit too early anyway so I don't mind waiting if you're willing to spend some of your precious time on it. >>11710 Once I'm back home I'll give it a spin and see if my autismo will break it or not.

>>11714 Hope your computer didn't catch fire lmao If you're having trouble usually it's easiest to install a fresh python environment or you have the wrong/no CUDA version installed. Crashes with no error are usually caused by OOM or your cpu/gpu overloading. OOM crashes are obviously also caused by OOM. Anyways, back to what I was originally posting about- I've been keeping an eye on Forge and there are at least three new code changes related to Flux and LORA optimization. Looks promising, though I think I'll wait a while before trying it again. Either way, kudos to the Forge dev for being so on top of things. Used like 80gb of virtual memory excluding my VRAM, insane. I think the one time I even managed to get a lora working it didn't work with anime, so I'm gonna wait for the tech to develop. Really hope someone makes that nsfw flux model.

>>11750 Note, forge is still worth installing because I think you can get some pretty satisfying results with Pony Diffusion (SDXL) and loras.

Continuing on this, I've found that not only can you "ehance" a picture but you can also not only "vibe transfer" but downright "image to image", it will grab whatever you give it and following what you gave it, depending on many factors will make a copy of it following your specific prompts. This lead me into a rather, rather deep rabbit hole, I'm merely emerging from it.

>>11778 From Pick1 to Pick 2, not bad at all! >>11750 >>11751 Alas, I still broke it... I'll keep screwing around with this until Flux is ready and/or I found a way (or bother my grease monkey friend) made by someone that knows how to wrangle tards with way too much compute power for their talent. I'm kind of seething that I went with 32 gigs of ram instead of 96, but I could know months ago that I would lean into AI stuff and JRPG/rep&pey stuff before hand, oh well. I'll monitor this thread with upmost attention.

>>11779 Not bad. I've kind of strayed away from img2img because I couldn't get reliable results but maybe I'll revisit it. I had better luck with controlnet but it seems unstable rn after all the rewrites. People are complaining of problems with it. I'm still beating myself up over not getting 64GB of ram yet I also can't justify buying more to myself. If I do then I'd just have to replace 4 entire ram sticks if I ever wanted anything beyond 64GB.

>>11791 Also, is there any specific way it's breaking or errors it's giving? Forge uses an absurd amount of RAM nowadays. I just configure windows to use a 50GB pagefile on one of my ssds and that kept it from crashing for me.

>>11791 It's all about the strength slider and perhaps screwing a bit with the prompts, I made another thread where I'm explaining as best as I can how to screw around with prompt, sooner or later I will have to make a list of all the tricks... someone called me out >>11782 and I explained that this thread is more about the tech monkey/hardwarepart, mine was more about prompting. >>11792 Nope, no error message, no crash, just simply vanish, it's not even in the process being stuck or anything, I don't know anon, I tend to be curse with things like that, perhaps because I'm using an heavily modified version of Window 10 (Atlas OS) there's probably some things that are screwed the wrong way (I'm mostly a gamer and I like to play those nice but horrendously optimised games)

>>11794 I don't really think I specifically intended for it to be about technical knowledge and news when I first created the last two threads. It just sort of happened that way because not many anons are active here, and my autism was left to frolic freely. Pretty sure it's like at most 4 people, and half the posts are me. Still, I do think it's an alright idea to have a thread specifically for prompts because of how cluttered this one gets with troubleshooting, news, and other nonsense. Though it's pretty much impossible to troubleshoot anything over an imageboard so I might create a server on a discord alt at some point or something. As for using AtlasOS- I would say you are very, *very* slightly completely and totally out of the standard support range for AI. It could be unrelated to AtlasOS but yeah that definitely doesn't help. Not sure what you quite mean by vanish. If the console has any text it might give clues about what the error is. It might tell you that it fails to generate, I have no idea.

https://civitai.com/models/556479/mge-monster-girls?modelVersionId=706589 Thought this was enough to warrant its own post to the thread. A LORA on CivitAI that adds support for 30+ MGE characters to PonyXL. Kudos as always to the creator.

I spent some time on spicychat and made two bots with the following premises: Smug, confident ancient mummy princess traps you, the explorer anon, in her tomb. I also made a clumsy stalker marriage-obsessed elf who always fucks up when she uses magic, either sexily or comically. I like them both, but smug mummy is definitely my favorite. Smug mummy: https://spicychat.ai/chat/76d9a081-d4c7-4dcf-bdfe-8f356435745e Clumsy stalker elf: https://spicychat.ai/chat/c07db718-9a78-45b6-aa39-9a6b52f158f4

>>11834 One day someone gotta hook up those lewd bots to those interactive sex toys, like onaholes with exposed APIs

>>11835 someday, bro. I've made so many smug mummies the last couple days. Who knows where AI will go in the next decade? What a time to be alive.

>>11834 Had a quick chat with both. Nepta jumped straight to paizuri which was pretty swell. I don't know if you wrote it into the definition, but the first thing she did when she got a hold of my rod was wrap it in bandages, which I also enjoyed. Overall solid bot. I led the elf on for a bit before shutting her down and decided to leave it at that because it was funny, so I can't comment much on her. She's otherwise very persistent, though. Anyways, you should also consider uploading them to Character Hub (Chub), I find its free service generally makes for better conversations and users can just use their own local installs with it if they want. Spicychat's pricing is completely insane.

>>11837 Damn i had no idea about Chub, thanks man i'll look into it tomorrow night. And yeah, the bandages stuff is in the definition, but the way she uses them sometimes surprises me. I also wrote in that she thinks she invented the concept of paizuri, lol. I've been adding some non-lewd stuff slowly as well. The elf still needs work, the magic she uses doesn't always fuck up as i would like, but i like her otherwise.

>>11837 I second this, Chub is great and it also provides download files for characterts with local text models like ggufs on SillyTavern.

Thanks to the power of S L O P P A, I finally managed to make my dream come true. Behold: Slimias! I guess the results could be better because I still have no idea what I'm doing, but it's a lot of fun to fuck with.

>>12078 Yup, aisloppa is just perfect for slimes, unsurprisingly.

>>12079 Yeah, but goop can only go so far when hiding deformities, especially when introducing non-slime things into the mix. Anyways, here are a few more with some more complex concepts, some more successful than others.