/mg/ - AI Monster Girls Thread

/mg/ - Monster Girls

Main monster girl board

Mode: Reply

Name
Subject
Message	Max message length: 20000
Files	Drag files to upload or click here to select them Max file size: 30.00 MB Max files: 5
More
E-mail
Password	(used to delete files and postings)
Misc

Remember to follow the rules

[ Return / Catalog / Bottom ]

AI Monster Girls Thread Anonymous 07/11/2024 (Thu) 11:26:06 No. 11282

I've been taking a look at AI again recently and the technology is very impressive. Seems most decent images these days come from an SDXL model called Pony Diffusion. https://civitai.com/models/257749/pony-diffusion-v6-xl Yes, it has furry shit but you can turn it all off since it comes with built in negative prompts for entire parts of the dataset. Despite being named after ponies 50% of the dataset was anime, and it's apparently the best or one of the best local models for generating it right now. It intuitively knows how to do sex, blowjobs, footjobs, etc without any loras. >What is a lora? Just think of it as an addon that adds additional features to a model it is trained for. (You can't use loras that aren't compatible with your model). I use Automatic1111 which is buggy as shit but has a lot of good extensions and I've put far too much time into it. https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-NVidia-GPUs Honestly, there are a few extensions that basically force me to stay now, and let me tell you about them: https://github.com/zixaphir/Stable-Diffusion-Webui-Civitai-Helper This labels everything you download from CivitAI and allows you to load in the "best prompt" provided with the lora, along with inserting all the model's trigger words and giving display images to each of your loras. https://github.com/Mikubill/sd-webui-controlnet Lets you use an image as direct reference for generating, so you could maybe even get tricky stuff like tail pussys or slime girls to work if you do it right. I'm not quite familiar with it yet but it's already performed really well. I had the best experience with anime canny and lineart models. You need to download separate models for it and put them in your a1111 folder's models/ControlNet folder. For SDXL, you can find those Here: https://civitai.com/models/136070?modelVersionId=267516 And here: https://huggingface.co/kohya-ss/controlnet-lllite/tree/main Finally, by far the most important one, ADetailer. https://github.com/Bing-su/adetailer It automatically detects faces in your picture and inpaints them to keep them from looking terrible. You can provide custom prompts for finetuning. Things to note: I recommend at least having a minimum of 6GB vram and an RTX 20 series or greater gpu. If you have below 9GB of vram, you should probably use xformers. Look at the automatic1111 wiki for how to do that. You should also check out https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Optimizations if it runs too slowly.

Anonymous 06/28/2025 (Sat) 21:46:55 No. 27948

>>27930 Running into an issue some other people are having where switching the checkpoint/model after booting can cause forge to generate only black squares or distort the model output (same seed generates a different image than before, usually with a different style or worse quality) It also might just trigger randomly, but usually I can generate 16-32 images so it's not intolerable. It's annoying, requires a full relaunch. I also just question whether or not the performance is even as good as it should be for my card, though I could just be imagining it. Seems almost slower in some aspects, but it could be that I've just adjusted to the new speed. I tested SwarmUI, it apparently uses Comfy as a backend, fortunately, it does let you configure the model folder so that wasn't an issue. Install was seamless. I think it's a bit slower than Forge for me, though. Something like 1.5x to 2x faster. Not having the same issue I had with Forge though since Comfy already has official support for 50 series.

Anonymous 06/28/2025 (Sat) 22:55:17 No. 27952

>>27853 It's not different as in a new function, it still works the same. A mix of vibe transfer, inpainting, and prose prompting. The 4.5 is just a little better at posing, object and background generation, and direction. Im still figuring it out though. the big update was 4.5 finally fucking getting vibe transfer, which is the NAI Mini-Lora imitation where you can upload reference images and scale their info extracted/strength individually.

Anonymous 06/29/2025 (Sun) 14:15:18 No. 28030

>>27574 >-Must have 16GB VRAM. tfw cucked by 12GB 5070

Anonymous 07/03/2025 (Thu) 02:07:32 No. 28643

Anonymous 07/03/2025 (Thu) 12:23:13 No. 28676

Anonymous 07/04/2025 (Fri) 05:43:20 No. 28782

>>28676 >that sausage picrel >cakesucc Very nice. The cream splashes look a bit off, but the idea is nice. Overall these 2 pics is what I want to see more of: actually using the tool, not just generating another simple POV fuck scene with 0 composition. Keep it up.

Anonymous 07/04/2025 (Fri) 14:05:26 No. 28839

And I'm still using a model that's like a year out of date.

Anonymous 07/04/2025 (Fri) 15:25:30 No. 28856

Anonymous 07/04/2025 (Fri) 16:49:44 No. 28886

Started to mess around with NAI again and honestly I'm kinda here for it

Anonymous 07/04/2025 (Fri) 17:57:29 No. 28909

>>28886 Like a moth to the flame, more will come. >>28782 Out of every one of those, cakesucc was by far the hardest to do. I had to manually edit on the gloops and the inpaint would continuously not recognize them, so i'de low-strength roll them, edit slightly, do again, etc. I also tried to do crumbs of the cake but same problem. A lot of AI is like directing a highly skilled blind person. Another fun fact, the frog girl wasn't that bad, but the fucking sausage/hotdog was so annoying because AI sees hotdog, and is like, oh yeah lets drizzle mustard and ketchup all over that bitch every time no matter what. I ended up fusing corndog (because those are usually on sticks) and sausage to finally get a useable standin, lmao.

Anonymous 07/04/2025 (Fri) 17:58:29 No. 28910

>>28909 first time i did (you)'s wrong, second time i forgot image, galaxy brain levels of retard. here you go, lol.

Anonymous 07/04/2025 (Fri) 20:10:48 No. 28939

>>28886 wow, first pic is really something else

Anonymous 07/04/2025 (Fri) 22:11:51 No. 28943

>>28676 >>28909 You used NAI for that? >>28839 What model? Also year old sounds like brand new to me. >>28939 Lol, lmao even. Unless you're sarcastic.

Anonymous 07/04/2025 (Fri) 22:20:42 No. 28945

>>28943 >Unless you're sarcastic. It has the AIslop shine but this actually isn't a problem with Demon's signature blue skin. other than that it's very good, only other weirdness I can spot as a layman is her right hand that's gripping the nipple and it's too small. The womb tattoo, dick, pussy and nipples with their piercings are all perfectly normal and that's impressive

Anonymous 07/04/2025 (Fri) 22:38:41 No. 28948

>>28943 >Lol, lmao even. Unless you're sarcastic. Man, who pissed in your soup?

Anonymous 07/04/2025 (Fri) 22:39:50 No. 28949

>>28943 Yeah, i use NAI and MS paint for my gens.

Anonymous 07/04/2025 (Fri) 22:46:53 No. 28950

Also, have this one which is a retarded joke from a different thread/board with no context whatsoever.

Anonymous 07/04/2025 (Fri) 22:47:27 No. 28951

>apologies for the non-monster pics Turns out I do have some free tries with NovelAI. 30 to be exact. First I experimented with some gyaru loli that was recommended as an "inspiration". I found it hilarious how the free trial was censored, but also tried to obey the lewd prompts as well, resulting in... well, you see what. Then I tried generating my rabbit character... Then I looked at my art... And it left a pit in my stomach. A simple mediocre drawing takes me like 90 minutes, the computer can do so much better in mere seconds. Sure, it needs polishing, but after polishing you have a reliable, insanely fast way to create any drawing. Whereas by hand every single simple drawing takes literal hours. Sure, you can talk about "soul" or whatever, and I do think my drawings have it, but so do the gens, so what? What can justify over 1000x increase in speed, all the while ALSO increasing the quality? I never even made a single cent with drawing, yet I still feel replaceable now. I thought I got over AI, and accepted "no rules just tools", but seeing it firsthand be better than me at drawing my character still had an effect. Damn...

Anonymous 07/04/2025 (Fri) 22:55:19 No. 28953

>>28945 >>28948 >messy hands >generic ai style >1girl, pov >no composition Cmon now. The tools have evolved, we're past that. >>28949 >Yeah, i use NAI and MS paint for my gens. >MS paint That's what I'm talking about, actually using the tool! Though I feel obligated to shill Krita, its very good, even if you're not into drawing. Recently it also implemented some live ai drawing tools, they look like absolute black magic, tho you need a graphics card for that. I saw some dude run it with 6gb of VRAM, though I can't make any promises. https://www.youtube.com/watch?v=nkVjXjzs4Z0 https://www.youtube.com/watch?v=Ly6USRwTHe0

Anonymous 07/04/2025 (Fri) 23:23:53 No. 28955

>>28953 Yeah, that looks to be kind of what i do but all rolled into one program. I wonder if it can link up to a style type via reference images or something. I usually like to generate several images of a character in different poses and slap them on as reference images. 1. imagine scenario 1a. if i already have characters made, pick reference images 2. Generate images of scenario and modify prose until i get a mostly acceptable pose or lighting or whatever i am looking for 3. adjust prose and generate variations of the "close" image 4. Draw and inpaint to blend the drawing without losing coherency The below images are a human maid and a dark elf maid (i have genned an absolutely absurd number of these two in solo and duo pictures). These aren't perfect but they are an example of keeping a mostly coherent style and characters/proportions. If Krita can do that and i can keep the styles i want from the base images i gen on NAI, that'd be fuckin' cool. I'll look into that for sure.

Anonymous 07/04/2025 (Fri) 23:54:05 No. 28961

>>28955 You can do consistent styles and characters using loras, though if you're making absurd amounts of OCs then training them all would be a pain. I dunno about a "reference" function in local programs, though I saw one "pseudo lora" thing where you load up 6 images into comfy and through some...thing it uses them similar to a reference. I see you're using mostly the same style, so a single lora would suffice for that imo.

Anonymous 07/06/2025 (Sun) 06:14:54 No. 29139

>>28839 UOOOOOO ANRIETTA-SAMAAAA 😭 >>28951 Cute! That squid character reminds me of the H.O.U.W.A.R. hat from TF2

Anonymous 07/06/2025 (Sun) 22:55:57 No. 29401

Anonymous 07/08/2025 (Tue) 07:14:56 No. 29584

Thread OP anon here, finally made a guide. Not done yet, but there's already enough information to be extremely useful to someone setting up Forge. I'll be updating it with additional information sometime later. https://rentry.org/4kaut9vg

Anonymous 07/08/2025 (Tue) 21:09:00 No. 29643

>>29584 >recommending pony I'll admit, I never used it, but I saw in other threads anons laughing at guys still using pony for anime. Do what you will with this information. I did use this model though https://civitai.com/models/952556/n40naillous-xl?modelVersionId=1066462 and can vouch for it, its really good. So any other anons wondering what to use, this one's great. Did see other people recommend WAI-NSFW. Haven't used that one.

Anonymous 07/09/2025 (Wed) 02:40:41 No. 29688

>>29643 I thought about changing out Pony base for AutismMix SDXL or something, or another Pony finetune. Bottomline is that Pony and Illustrious based finetunes are the best *local* anime models. I think Pony is worse at some concepts without loras though, I might add a note about that.

Anonymous 07/10/2025 (Thu) 19:50:23 No. 30081

Anonymous 07/11/2025 (Fri) 15:08:31 No. 30249

More demon gyaru usin DEM HANDS

Anonymous 07/11/2025 (Fri) 17:28:00 No. 30295

Anonymous 07/11/2025 (Fri) 22:55:47 No. 30354

>>30249 How hard is it to gen multiple arms doing what you actually want them to do?

Anonymous 07/11/2025 (Fri) 23:55:17 No. 30373

>>30354 It likes to gen more than the four requested, and it gets confused with multiple characters. Solo? It's not that bad. Two people? you'll need to inpaint.

Anonymous 07/12/2025 (Sat) 23:11:17 No. 30563

I FINALLY GOT A DECENT PC WITH A 5060 TI 16GB! AND IVE BEEN FIGHTING WITH DRIVERS ON LINUX MINT ALL DAY I'm close to switching to shitdows, but I want comfortable local stuff, not booting to a different os whenever I want to gen something....

Anonymous 07/13/2025 (Sun) 05:36:09 No. 30618

I put together a quick, low-effort Dullahan bot to see if Deepchink could keep up with the concept. It was hit or miss most of the time, but I can't stop thinking about this particular exchange.

Anonymous 07/13/2025 (Sun) 17:24:57 No. 30717

>>30563 >I'm close to switching to shitdows Nope, can't do even that, the old drive with windows is screwed with some proprietary tiny bullshit screws, so I can't unscrew it, and for some reason I can't install from usb because of AMD cpu. So I'm fucked because nvidia on linux, and can't install shitdows because of AMD cpu that needs some drivers. Fucking hell, I'm mad.

Anonymous 07/14/2025 (Mon) 14:58:50 No. 30959

>>30717 Finally managed to get it working, maybe the windows ISO was fucked or something, I dunno, I doncare. First ever local generation on my own pc. I went overboard on purpose, tried 3333x3333 pixels with 33 steps. Took 7 minutes, the model went stupid and did what you see, but I kinda like it. Others done at normal resolutions returned normal generations. nnaillousXlSeries_n40naillousv2Xl model already knows many artists you can find on gelbooru and alike, so I had some fun with that as well (tried Kenkou Cross, machigami_yoh, jin_(mugenjin) (author of Miss Black General), harano_kaguyama). Also used Dot_Comics-000018 lora, it gives a nice monochrome manga style. the future is now old man, and it has massive ai tits Anyway, I'm feeling nice, throw requests you want to see (only sfw though).

Anonymous 07/15/2025 (Tue) 05:10:45 No. 31119

Anonymous 07/15/2025 (Tue) 05:49:27 No. 31121

Requesting a Lamia wearing a sombrero and a poncho. Bonus points if she has a comically large fake mustache attached.

Anonymous 07/15/2025 (Tue) 16:16:09 No. 31179

>>31121 >a lamia How about 240+ lamias? Man, this was fun. Also learned a lot: nnaillousXlSeries_n40naillousv2Xl (N40 for short) knows a ton of artists by default. Not adding any artist tags usually resulted in trash (near ms paint levels, look at first gens in the collage). Adding "masterpiece, best quality" as first 2 tags makes the end result a little better, but using artist tags, or something akin to them seems like a much better choice. Changing tags and queuing generations faster than your gpu can do them seems like the optimal way to use the tool (I've yet to get this month's electricity bill though, haha). After you get a ton you just sort through them, upscale/edit the good ones. Its like photography- first you spam the shutter, then you pick the best. If you want I can upscale some pics from this set (ie raise quality, fix fucky faces/hands/details), just show which one in the collage and I'll try.

Anonymous 07/15/2025 (Tue) 16:25:27 No. 31183

>>31179 Some more highlights.

Anonymous 07/15/2025 (Tue) 20:43:15 No. 31231

>>31121 Based request. I thought it would be apropo to make her a rattlesnake, and i like the other anon's idea of making her shred on a guitar.

Anonymous 07/15/2025 (Tue) 23:03:55 No. 31256

>>31231 >Based That's not how you use this word. >rattlesnake Good idea though.

Anonymous 07/16/2025 (Wed) 01:45:32 No. 31297

>>31179 >>31183 >>31231 Very good, thanks hermanos. I'm curious about that sequence mixing a bunch of different anime characters or distinct artstyles with the request, did those come about naturally or did you put them there yourself? >Rattlesnake I actually considered suggesting that myself, but I didn't think there'd be enough for the model to pick it up.

Anonymous 07/16/2025 (Wed) 03:52:18 No. 31324

man four arms really does offer up a lot of unique posing potential.

Anonymous 07/16/2025 (Wed) 09:44:10 No. 31361

>>31297 I mentioned in the first post I was experimenting with artist tags. The only ones I explicitly prompted were Tomoko, Konata, Ranni and "blonde_hair, long_hair". Stocking came out when trying to force it to use the PSG artstyle. Asuka/Nadia came out when using the OptionalTypo artist tag (the guy draws them a lot, and has a gainax-ish style in general), the FLCL bitch came out when using sadamoto_yoshiyuki (official artist for many core gainax things). I also tried making one in the style of van gogh's starry night, but the model thought I wanted fate's van gogh character. For most generations I didin't put any details on the character so the model "decided" itself what hair colours, styles, faces, tail colours, etc to use.

Anonymous 07/16/2025 (Wed) 09:47:56 No. 31362

>>31361 >>31297 >>31179 Also worth noting is that that collage has everything I generated then, in chronological order starting from top left. So you might also see when I added new tags or refined the idea (sky, sun, dancing, perspective, dynamic, 2 girls, etc).

Anonymous 07/16/2025 (Wed) 10:03:21 No. 31363

>>31324 Its like having 2 characters posing at once. I wonder if something similar could be achieved with 4 eyes...

Anonymous 07/16/2025 (Wed) 12:54:48 No. 31382

https://github.com/lllyasviel/FramePack?tab=readme-ov-file https://github.com/lllyasviel/FramePack/discussions/53 Finally managed to get it running on my pc with 5060 ti16gb. It didn't work initially, had to also follow the guide in the second link for the 50 series cards, then (eventually) it finally worked. The results are definitely slop-ish, but the potential is there! The 10 sec one took around 50 minutes, the shorter one around 20 (I think, didn't measure it unfortunately). lllyasviel says his 4090 does around 1 frame in 1.5-2.5 seconds. So relative performance appears to be more or less alright, but I'd have to measure it better. Either way, just check your GPU relative to his and mine and you should get a rough idea what generation times you can expect. Also of note, all my 32gb of RAM were filled during generation, but not all VRAM according to task manager. Websites were definitely deleted from ram because after clicking on a few tabs they had to be reloaded. Do what you will with this information. Reading into it further (https://github.com/lllyasviel/FramePack/issues/151#issuecomment-2817054649) seems like RAM speeds and SSD/HDD speeds also have non negligible effects. Hey 5090 anon, I have a request, try running it on your machine and post results and generation times, pretty please? Do whatever you want, I only want to know how long it takes on the highest end gpu.

Anonymous 07/17/2025 (Thu) 18:52:48 No. 31647

>>30739 >Need more amazonian Gazers. <huh, lets just check if N40 can do Gazers without loras, nothing much <201 pics later I can't wait for the electricity bill! Anyway, experimented again, this time with very low step counts to save time (and also electricity, lul). Turns out you can go VERY low, at just 10 steps you start getting some ok stuff. Going above 20 seems pointless from my testing, 15 would be the fast option for spamming images, then just upscale the good ones (which I still didn't do, heh...). Aside for that, tested artists again. I find it funny how vanripper makes them more clothed.

Anonymous 07/22/2025 (Tue) 21:44:47 No. 33064