No.186
To "create" the OP image I used the following prompt and you may notice that many of them were ignored. That's just how it goes.
(masterpiece), (highly detailed:1.2), absurdres, HDR, beautiful lighting, explict, nsfw, sharp focus, cat ears, (megumin:1.3), red eyes, eyepatch, wizard hat, leather gloves, red cape, (huge penis:1.3), delicate, nipples, focused on waist, belly, solo, perfect face, beautiful, shiny skin, sweat, from side, singing, [squatting:straddling:squatting:.3], (riding anal dildo:1.3) beautiful detailed concert stage, explosion in background, spotlight, audience, (realistic shadows:0.6)
Negative prompt:
(bad_prompt2:0.8), 3D, lowres, navel, feet, legs, extra nipple, bad anatomy, feet, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, (deformed), [blurry], disfigured, poorly drawn face, mutation, mutated, (extra_limb), (ugly), (poorly drawn hands), messy drawing, pussy
Steps: 34, Sampler: Euler a, CFG scale: 8.5, Seed: 3688624663, Size: 576x768, Model hash: d86b2baa, Clip skip: 2, ENSD: 31337
"explict" is a purposeful typo by me because when someone made the furry model he mistyped "explicit". Why is the furry model involved with this Megumin, you ask? Because the furry stuff is loaded with much more sexual content so penetration and anuses actually show up. You can even have penises viewed from behind, which is what I was going for, but the "AI" isn't perfect.
So, what I did is I took the furry model and merged it with the Chinese 'Anything' model to pick up some of its anime-ness including Megumin's existence. The downside is that her face would look horrendous since she's not an animal, but you can't see it from this angle.
Attached to this post are the embeds I have. Extract the ZIP into your main AI folder (keeping the embeddings folder) and you can call them by name. Of particular interest is the "bad_prompt" and "bad_prompt2" ones which are meant to be put into the negative prompt at 80% strength to try and automate the negative prompt process. I should probably try experimenting with all the other text in the negative prompt gone.
But as I said here
>>>/qa/100161 people otherwise don't make embeds any more in favor of hypernetworks
No.187
>>186>(highly detailed:1.2)>[squatting:straddling:squatting:.3]What do these bits with colons and numbers do? Is that a way of setting a precise strength or something else?
No.189
And here's an example of that prompt in action, including the same seed.
>>187I'm not entirely sure how stuff is "read" by the AI, but...
>(highly detailed:1.2)It's in parenthesis so 1.2 is intensifying it by 20%. I've seen people use parenthesis in different ways, but I'm not sure how that works, if it does. Many people like myself are just learning as we go.
>[squatting:straddling:squatting:.3]It's supposed to be drawing one, then the other, and then the other. It only does this if it's in brackets. I'm still not sure how this one works. I mentioned it in that /qa/ thread with the angel/demon thing
>>>/qa/100156 because it's really obvious when it's something like that, but for something like this I don't even know if it does anything. My intent, at least, was for the AI processing to start drawing her squatting down but also in a similar position to straddling something. Again with the angel/demon example you can see how that imageset in the middle was a cross between angel Sachiko and demon Sachiko so I was going for something like that.
No.190
>>189By contrast, here is the prompt in the default 'Anything' checkpoint model thingie. While it is certainly of a better style and quality, you would need to produce dozens upon dozens of images before you might get one that had a dildo actually penetrating her, and it would probably look weird. I've yet to find the magical recipe of combining checkpoints to have good faces while also having sex stuff.
Also, the "Anything" model is terrible at putting penises onto girls and often even boys. You need to give it an intensity of like 1.8 to make one appear, but when it's that high it's rarely one of them and they're all usually deformed.
No.191
>>186What's the CFG thing anyways?
No.192
>>191You can hover most things to get a tooltip. The CFG controls how much it follows your prompt input. 1 would be something closer to completely random and results in a garbled mess. The NAI model is built around 12, so I generally hover around there, although maybe I should start increasing it since I'm getting better at writing the prompts.
(masterpiece), (highly detailed:1.2), HDR, beautiful lighting, nsfw, (minase iori:1.3), (loli:1.1), (chibi:1.3), sharp focus, (huge penis:1.2), nipples, focused on waist, solo, perfect face, beautiful, shiny skin, sweat, (pink lingerie), pink gloves, pink bunny ears, looking at viewer, looking down, first person view, from below, blushing, angry, claw pose, bedroom
Negative prompt: (bad_prompt2:0.8), lowres, navel, feet, legs, extra nipple, bad anatomy, feet, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, (deformed), [blurry], disfigured, poorly drawn face, mutation, mutated, (extra_limb), (ugly), (poorly drawn hands), messy drawing, pussy
Steps: 34, Sampler: Euler a, CFG scale: 4.0, Seed: 2215924190, Size: 448x576, Model hash: 579c005f, Hypernet: Momoko-hn2, Clip skip: 2, ENSD: 31337
No.194
Well here's my first attempt at generating a Me-goo-min.
Prompt: {{{{loli}}}}, {{1girl}}, impressionism, expressionism, , , , , , , , watercolor (medium), cum, pov, {{{megumin}}}, ahegao, orgasm, upper body, bare shoulders, collar, nipples, [choker], {hat}, symmetry, blush, arms, red eyes, , squinting, cum, dramatic shadows, {{cum in mouth}}, cum on face, outdoors, loli, symmetrical mouth, cum on hair, open mouth
Negative prompt (plus NovelAI defaults): huge ass, doujinshi, comic, huge breasts, yuri, bowtie, bound wrists, penis, nipples, tongue
Steps: 28
Scale: 6.33
Seed: 309519879
Size: 512x768, but then upscaled and cropped
Sampler: Euler a
Model: NovelAI Diffusion Anime (Full)
No.195
>>194What are the brackets for?
No.196
>>195I don't know the exact details, but as I understand it curly brackets strengthen the input whereas square brackets weaken it.
No.197
NAI itself has different tagging systems. The offline version people like myself use added the decimals to replace the multiple brackets and such. Well, both work actually, but numbers seem a lot easier to manage since it's some weird multiplier with them. I think each pair is a 1.1 multiplier. NAI uses {} for emphasis whereas the offline uses (), but both use [] for less emphasis (or you can use a number below 1 like 0.7 for offline version)
No.202
Some info on the "models" (also called checkpoints).
When I say "merge", it's when you use the Checkpoint Merge tab in the WebUI to, well, merge two different models. I don't know the way it works, but some stuff is sacrificed from the main model when you add a new one to it. I.E if you have a 2D one and you add the "gape" model which was trained on RL sex stuff you'll get better results for that 2D versions of that sex stuff, but you'll find it's worse as doing specific characters or something. I'm only including models people I've personally messed with or merged.
Stable Diffusion (SD) - The base one. Super generic and good at real life stuff, but not anything related to 2D escapism. Sometimes used as a merge with the others to add more variety to them
Waifu Diffusion (WD) - One of if not the first 2D models. Atrocious name, but it started to show the potential of this for 2D. Not great, but apparently a new version is in development and will be released around the 1st which could be major
NovelAI (NAI) - A paid service with its own model. It was thankfully leaked almost immediately. It scraped danbooru and proudly told people about it which lead to a bunch of drama and takedowns on danbooru itself. Fuck these guys. Nice model that obeys danbooru tags, although it seems like it couldn't scrape the paid tags like loli. Because of that, this model and similar ones respond much better to "chibi" or SD descriptions like "12 year old girl". Very common as the base model in merges
Anything (also called the Chinese model) - Another scrape of danbooru. I don't know what they did exactly, but it generally results in much higher quality art than NAI. However, as I said earlier it's absolutely horrendous at penises
Trinart - A scrape of modern 2D art (including Japanese) that is apparently fine tuned towards quality, however it does not use danbooru tags and does NOT include NSFW stuff. Used as a merge to enhance others sometimes
Rule 34 (R34) - Scrape of a rule34 site similar to the way the others scraped danbooru. I never had luck when I messed with this, but I never gave it much of a chance, either. Seems to produce a lot of bizarre monstrosities that the other models don't. It has its own tagging system and I don't want to learn a bunch of them. I've seen merges include it, but I don't know what exactly it adds to them.
Hentai Diffusion - Some combination of danbooru and R34 scraping. I can't remember much about it because I never went back after the NAI leak.
Yiffy and Furry - Two different anthropomorphic ("furry") models trained on e621 the furry booru site. You could go to the /trash/ board on 4chan and search for the stable diffusion thread to see what they make with it. I use this as a merge into other models because it's significantly better at the crotch region. I don't know if it's because of Japanese censoring on danbooru or if it's just that Western artists generally add much more detail. This comes at the price of face detail since I make humans, but hypernetworks can cancel this somewhat.
Gape - Used as a merge to make the others better at penetration, gaping orifices or those holes down there in general. I'm not sure it's worth it, but I've downloaded it and merged it when following guides. Now that I've thought about this probably isn't something I make use of personally since I use the furry model merged into stuff.
No.203
>>202>Anything (also called the Chinese model)Was reading some Github posts about this model where someone was speculating that it may actually itself be another leak from NovelAI since the same image with the same hash could be created as with AnythingV3. Although I'm not sure if that's really true or not.
No.205
>>202>NovelAI (NAI) - although it seems like it couldn't scrape the paid tags like loli. No? NovelAI works fine with the paid tags - the first two images here are a comparison of generations (with the same seed and basic settings) for
1boy and
1boy, shota. It's also worth mentioning that NovelAI has its own furry model too, although I don't know if it is the same as either of the ones you mention - the new three images here are respectively for
male, human,
1boy, human and
1boy, cub, human in the furry model (it also works with shota but the result was a bit too far in the shota hentai direction for Kissu).
No.206
>>205Hmm, strange. I never had luck with those blacklisted tags myself. I made a bunch of comparison images for myself, but already deleted them weeks ago...
No.207
>>206Huh, I'm not sure why that would be. I'd suggest you were using the censored NovelAI model, but even that seems to work with shota (although for some reason generating males is much more difficult, which is why I didn't bother including the results in that comparison image). I also have test loli/shota images that I generated over a month ago using the leaked model, so it's definitely not a new change either.
That being said though, I have noticed that for some reason the rape tag doesn't work on NovelAI (although it is easy enough to generate such content regardless), even though that one
isn't premium-only on Danbooru, so there does seem to be a certain degree of prompt censoring at some level.
No.208
I've been spending so much time generating images to try and find a good seed, and then I realized I can just cut down the amount of images I generate to like 3 to see if I like the process and then I can go generate a full set once I find a seed I like....
No.209
So, I've just spent the last couple hours learning the ins and outs of this X/Y plot prompt S/R thing, and while doing it I realized that you don't need to stick to one scale that changes with a value over time. You can actually substitute out the bit you select with anything and it'll compile. This is very useful for getting the right images in a selection, or trying to make a more gradual shift process instead of leaving all the hard work up to the AI itself leading to many misinterpretations.
For this image I used:
Positive prompts
(masterpiece), {{ishikei}}, (highly detailed:1), 1girl, 1boy, masterpiece, {nipple}, {straddling}, {pink hair}, [flat chest loli ahegao: surprised breasts gyaru with cum on body sweat sex:0.9]
Negative prompts
(bad_prompt2:0.8), 3D, lowres, extra nipple, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, (deformed), [blurry], disfigured, poorly drawn face, mutation, mutated, (extra_limb), (ugly), (poorly drawn hands), messy drawing
Sample steps: 40
Width: 512
Height: 640
CFG: 16
Seed: 2683737196
Then I used what follows in the image for my X/Y plot prompts. Maybe it's a bit too fetishistic, but I think it really does a great job to illustrate just how creative you can get with making changes to the image.
No.210
I know that before I've seen people making images in an artist's likeness, but how would you go about doing that? As a test I tried replicating Asanagi's style, but I just can't seem to generate anything that has that unique look to it.
No.211
>>210In an attempt at plausible deniability the AI people scrubbed all artist names from the data so they can't be called by the prompt. This isn't true for the furry models (why are the furries always the best at this stuff?), but anything built on the big danbooru scrapes don't have them.
So, the answer is hypernetworks and embeds. I already uploaded a zip of my embeds in the thread and one of them is asanagi. There's also asanagi hypernetworks but they're 78mb each so I can't just upload them here.
https://rentry.org/sdgoldmine#hypernetworks No.212
>>211and here's one of the asanagi hypernetworks in action. Even at 1.3 intensity it's weaker than the embed stylewise, but it does properly allow Takane to shine. I've been messing around with e621 tags to see which might work with Anything and the answer is a resounding "nothing". Danbooru uses 'muscular' instead of 'muscles' so it wasn't picked up. 'stocky' is e621's version of 'plump' and it was ignored, and so on. I forgot Anything doesn't do penises so that one was wasted. I don't think it matters if you use something like "pale skin" instead of "pale_skin", but I should test that.
The stuff inside the curly brackets are wildcards used by the dynamic prompt extension which you can find on the Extensions tab. You write a text file with words and it will insert them randomly. REALLY nice to have. It also has its own "magic prompt" system of wildcards that you can choose to enable to add even more variety.
(masterpiece), (highly detailed:1.2), (akihiko yoshida), HDR, beautiful lighting, (asanagi:1.3), nsfw, solo, gothic, pale_skin, (shijou takane:1.3), vampire, (veiny penis:1.06), sharp focus, gothic lolita, nipples, perfect face, (huge muscles), stocky, {__expression__}, {__pose__}, blushing, beautiful detailed castle, (realistic shadows:0.6), professionally retouched, vibrant red lighting
Negative prompt: (bad_prompt2:0.8), feet, mature, adult, multiple boys, lowres, group, vaginal, bad anatomy, out of frame, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, (((deformed))), [blurry], bad anatomy, disfigured, poorly drawn face, mutation, mutated, (extra_limb), (ugly), (poorly drawn hands), messy drawing, pussy, censored
Steps: 38, Sampler: Euler a, CFG scale: 12, Seed: 3120914168, Size: 448x704, Model hash: 6569e224, Hypernet: asanagi, Clip skip: 2, ENSD: 31337 No.234
Okay, I organized them.
This is the effect of 146 artists on the Chinese
Anything Full EMA model which you can find a torrent of here:
https://rentry.org/sdmodels#anything-v30-38c1ebe3-1a7df6b8-6569e224. I organized them into minor, medium and major change:
Minor Change has a minor effect, but still an effect. Some of them might even just be considered "noise", but I did notice some patterns like larger heads, a more mature face or bigger bellies so I can't count them out. If AI was an artist I'd believe that these were all from the same person.
Medium Change makes it look as if it were drawn by a different artist, but it's still what you'd expect to see on a booru on an average day. Some might look similar enough that you'd accept it as a seed variation at first glance, but there was enough of a difference that I think it was actually influenced by the tag.
Major Change has a huge influence on how it looks. I go by immediate impression and most of these you would notice from the thumbnails and I included a screenshot to demonstrate.
The images:
https://mega.nz/file/sl8glAjJ#ZxcpkF6qqrW8i5_VOx9Fuv283m-XI7OfFLx9bE4Ke1QNote that I did these at
1.42 intensity to make sure that they would heavily influence the resulting image. You can easily tune them down for a more sensible result, as some of the Major ones are incomprehensible. In hindsight I should have included a muscle tag because artists can heavily influence those, but it's something to do in the future I guess. I should also do a set of these with barely any descriptions so that the artist tag itself influences it more. Something like "nsfw, 1girl, beautiful background" or something.
Feel free to share the RAR around, I included a kissu.moe link inside.
No.235
Note that since the images were not directly uploaded on an imageboard the extra data was not purged and you open the PNG Info tab and drag and drop to get the data used to produce it. I.E when you drag and drop the _ORIGINAL_ image this is what you'll see. But if you want to try it on NAI or something here is the prompt:
nsfw, (penis:1.2), (solo), 1girl, (fins), (scales), feathered wings, plump, belly, long purple hair, jewelry, fish, cute face, warm smile, nipples, beautiful detailed water, wading, splashing, yellow mountainous horizon, pink waterfall, bubbles, rainbow, starry sky
Negative prompt: (bad_promp2:0.75), (3d:0.9), feet, lowres, fleshpile, bad anatomy, worst quality, pussy, censored
Steps: 37, Sampler: Euler a, CFG scale: 12, Seed: 1774960972, Size: 576x768, Model hash: 6569e224, Clip skip: 2, ENSD: 31337
No.236
>>235Oh, and this reminds me. Put "3D" in the negative prompts. It's a meta tag on boorus and selecting against it will give you things that look less generic because they're not influenced by all the 3DCG stuff that people never give proper shaders to, although I'm bit surprised danbooru has enough of that low quality stuff for it to get influenced.
Lisa Frank at 1.42, George B. Bridgman at 1.2.
If only penises weren't so prone to messing up...
No.237
(deleted and reposted because I deleted the remaining "_" symbol in each image filename so now you can just copy paste the names)Okay, I organized them.
This is the effect of 146 artists on the
Chinese Anything Full EMA model which you can find a torrent of here:
https://rentry.org/sdmodels#anything-v30-38c1ebe3-1a7df6b8-6569e224. I organized them into minor, medium and major change:
Minor Change has a minor effect, but still an effect. Some of them might even just be considered "noise", but I did notice some patterns like larger heads, a more mature face or bigger bellies so I can't count them out. If AI was an artist I'd believe that these were all from the same person.
Medium Change makes it look as if it were drawn by a different artist, usually, but it's still what you'd expect to see on a booru on an average day. Some might look similar enough that you'd accept it as a seed variation at first glance, but there was enough of a difference that I think it was actually influenced by the tag.
Major Change has a huge influence on how it looks. I go by immediate impression and most of these you would notice from the thumbnails and I included a screenshot to demonstrate.
The images:
https://mega.nz/file/VgFBkYJR#65ospjxqcT20V0gwdoHuFzpp6SngFWO2AAjeWHS91o8Note that I did these at
1.42 intensity to make sure that they would heavily influence the resulting image. You can easily tune them down for a more sensible result, as some of the Major ones are incomprehensible. In hindsight I should have included a muscle tag because artists can heavily influence those, but it's something to do in the future I guess. I should also do a set of these with barely any descriptions so that the artist tag itself influences it more. Something like "nsfw, 1girl, beautiful background" or something.
Feel free to share the RAR around, I included a kissu.moe link inside.
No.272
oh yeah, and if you want to try the list yourself:
pino daenim, yoshiyuki yomino, alphonse mucha, shohei otomo, heywood hardy, Albert Bierstadt, john berkey, michael & inessa garmash, lisa frank, Albrecht Dürer, andrew wyeth, camille pissarro, camille corot, anna mary robertson moses, charles adams, charles dana gibson, coles phillips, dale chihuly, dean ellis, dorothea tanning, edmund dulac, Édouard Manet, Edward Burne-Jones, Edward Hopper, Gaston Bussière, George Inness, Gustaf Tenggren, Gustave Moreau, Hieronymus Bosch, Ivan Aivazovsky, John Bauer, John Constable, Joseph Cornell, Kandinsky, Kate Greenaway, Kelly Mckernan, Louis Glackens, Louis Janmot, Makoto Shinkai, Malevich, Marianne North, Mati Klarwein, Maxfield Parrish, Mike Mignola, Milton Caniff, Odd Nerdrum, Odilon Redon, Paul Delvaux, Paul Lehr, Pierre Bonnard, Raphael Lacoste, Rene Magritte, Richard Dadd, Salvador Dalí, Seb McKinnon, Tamara Lempicka, Tsutomu Nihei, Wassily Kandinsky, Alberto Vargas, Aleksi Briclot, Akihiko Yoshida, Alex Katz, Alex Toth, Alexander Jansson, Alex Horley-Orlandelli, Anna Dittmann, Antonio Mancini, April Gornik, Arnold Böcklin, Art Spiegelman, Artemisia Gentileschi, Arthur Rackham, Asher Brown Durand, Austin Briggs, Ayami Kojima, Bastien Lecouffe-Deharme, Beatrix Potter, Brian Bolland, Brian Froud, Bruce Pennington, Canaletto, Clara Peeters, Clyde Caldwell, Coby Whitmore, Damien Hirst, Daniel Merriam, David Firth, Diane Dillon, Donato Giancola, Dr. Seuss, disney, E.H. Shepard, Earl Norem, El Greco, Elizabeth Shippen Green, Emmanuel Shiu, Esao Andrews, Eugène Delacroix, Federico Pelat, Frederick McCubbin, Gediminas Pranckevicius, Geof Darrow, George B. Bridgman, Gerald Brom, George Cruikshank, H.R. Giger, Harvey Kurtzman, Hokusai, Inyi Han, Johfra Bosschart, Jon Whitcomb, Junji Ito, Kim Jung Gi, Laura Muntz Lyall, Laurel Burch, Laurie Lipton, Lee Madgwick, Liam Wong, Lise Deharme, Luis Royo, Lynda Benglis, Lyubov Popova, Maciej Kuciara, Peter Max, Peter Mohrbacher, Pieter Claesz, Ray Caesar, Rebecca Guay, Richter, Rob Gonsalves, Robert McCall, Robert McGinnis, Robert Rauschenberg, Rockwell Kent, Rodney Matthews, Ross Tran, Sir John Tenniel, Stephan Martiniere, Steve Dillon, Steven DaLuz, Sylvain Sarrailh, Takashi Murakami, Takato Yamamoto, Takeshi Obata, Tomer Hanuka, Wendy Froud, franz xaver winterhalter
No.308
Also... I AM NOW ACCEPTING TRAINING DATA FROM KISSU USERS! You need to give me images and then text files with booru tags that match it. A good way to get those is to use Grabber and follow the directions in this post:
https://desuarchive.org/g/thread/89739263/#89739456I personally used the "bulk rename utility I linked in the "cool software thread" on /qa/ to rename the text files.
I still had to do a bunch of cleaning, though, especially after cropping stuff out. I'm not sure how important that is, though. I guess I probably did overkill, but maybe that will give me great results?
So for instance if you wanted a new character, you'd get me like 10-50 images (or more?) of that character, properly tagged, and then I could built an embedding or hypernetwork out of it and stuff. COOL!
No.309
>>308This is the sort of thing that must work with 512x512 pictures, right?
Also, does it have to be all booru-style tags or can they be accompanied by other text? The Stable Diffusion model was trained on the LAION 5 billion image/text dataset which had natural language descriptions.
My understanding of training embeddings or hypernetworks is that they still work when you give them just images, and text descriptions are an optional extra.
No.310
(I'll answer that post after I made some blog posts)
>>308Training is going well, but I kind of regret going for an option I read was quite a bit slower, but more "safe". Maybe it would have been better to mess up and lose a night of progress than take 80 hours altogether. Well, this is something you can really only get a hang of by testing it yourself, but sadly this stuff is much larger timescale.
I have it set up to produce an image from random tags inside the batch of text files and watching it progress is really amazing.
Pic related is from 60 steps, from a process I set to be 20,000
No.313
>>312990
The body shape really does look like his style here, except perhaps a bit more muscular
No.314
>>3131400
I did crop a couple images so "face out of image" or whatever the danbooru tag was exists in these
No.315
>>314and I'll end it here with the 1870 image.
It's getting really good at doing the penises and body, but the faces not so much. I'm a bit worried if this means faces are going to be screwed up, or if I can chalk it down to me never using 'face' in the text files apart from the booru tags for smile and such.
You can fix faces with the inpainting thing, but I'd rather not
No.316
step 2020
The good news is that this training stuff doesn't seem as intensive as general image generation, although with so much VRAM reserved I don't think I could play a modern 3D game.
This image looks like a failure in that regard, however looking at its tags it did not include 'penis', just 'veiny penis' and 'huge penis', which I believe never exist on a booru without the 'penis' tag so it's possible that this image never attempted to draw it.
>>309From what I've read people say to use 512x512 unless you're some expert or have a really expensive card with 24gb of VRAM or something. This isn't actually written down anywhere from what I can tell, at least in layman's terms. I'm relying on 4chan and github comments.
It doesn't NEED to be booru tags, but I'm personally most interested in using tags and characters that you'd see on a booru and not an amalgamation of reality, so that was what I mentioned. "Huge breasts" has a very different meaning in 2D, for example, and of course stable diffusion's newer versions have begun to trim away nsfw and the new one (2.0) has done its best to proudly purge sexual stuff altogether.
But yes, SD's system is meant to have a more natural flow of words as you mentioned. The danbooru models are built upon it and function alongside its descriptions, but I specifically want to invoke tags and assumed others would as well. I guess I should clarify that my offer to train stuff does
NOT extend to real life porn or people. I don't want to look at that stuff and it's even more morally reprehensible than copying an artist's style to masturbate.
Training without tagging? I'm aware of it for the "aesthetic gradient" thing, however that may work, but I don't know of it. Everything I read about with hypernetworks says to include text files of the description. Well, there's the 'deepdanbooru' script that automatically detects them, but it's not perfect. I really do think the text files are necessary, at least for the booru stuff.
I think dreambooth is supposed to work without tagging, but that's a whole other ballgame and I have no idea how it works. Some people say it sucks, some people say it's amazing. It's 12gb VRAM minimum, which I do now meet, but getting information on it is extremely difficult. I don't really have faith that it'd recognize giant penises, either, when everything struggles so much with them even when you directly point them out.
No.317
>>316>This image looks like a failure in that regardOops, moved some sentences around. I was referring to what is now the last sentence of that post in regards to generating penises.
I'm going through 4chan search (sadly it resets every month) and desuarchive trying to find information and I'll catalog it when I'm done.
Here is an example of people specifically talking about embeds, but I don't really understand the last paragraph. Oh well.
No.318
I stopped at 3000 steps. I think the one I started one was doing faces better at 50 steps than the one I'm at 3000 steps on, so I'm reverted to the one that I briefly tried last night before switching to this slower method.
I've also noticed some tags aren't correct, as I grabbed some images from danbooru, gelbooru, r34, and sankaku and they aren't all shared. R34 in particular is too different, and some stuff is obsolete. I.E "cock-tail" used to be the name for a penis visible from behind, but at some point it was switched to "penis tucked".
Also, I think the penises I had in isolation and cropped are causing problems, so I'm removing those. I thought it'd be good for more data to learn from, but when it's isolated like that it's not contributing to where the penis is supposed to be located, so that must be bad information to include.
Still, I'm amazed at how well it's doing. As far as I'm aware SinenSian never drew Chun Li, but here she is. Also, I don't have the VAE active which is the colored are muted and it might look worse than it would in reality.
COOOOOOOOL!
If someone wants to mess with this incomplete and flawed one, here it is:
Rename it as 'sinensian.pt' and put it in your hypernetworks folder and remember to include 'sinensian' in the prompt!
I'm using litterbox because I'm going to replace it with a better version by the time this link expires in 3 days, and plus it's built on plagiarize an artist so...
https://litter.catbox.moe/6e6tb6.pt No.321
>>320It still have to wait and see what it's doing, but so far it seems to be learning REALLY fast.
This is after only 50 steps! Compare this one to the 60 and 300 in post
>>310 and
>>311 respectively.
No.323
>>321Bleh. As nice as the progression looked in this method, I must have done something wrong because it produces a lot of noise in actual use. I'm not sure why. I also learned that apparently you're not supposed to train on 'Anything' and instead use NovelAI, so maybe that was it.
However, the good news is that someone shared a model merge built on dickgirls and it works really well, so I think I'm going to forgo this sinensian thing for now. This merge is also obviously partly trained on him anyway.
Here it is: it's called chunkymix:
https://drive.google.com/file/d/1nGzWG0UIbvhHY0JCC0zM2N33Wg4OyDh8/viewYou should manually turn a VAE on when you switch to it or else things will come out grey and washed out. I use Anything's VAE with it. Follow the thing I mentioned here about adding those options to the main menu to make it a lot easier:
>>>/qa/100551 No.324
This is one of the best prompts ever. I built it another guy's generic thing to put Patchy in a restaurant with burgers, and if you set CFG low (7 for instance) you get lots of nonsensical stuff and it's amazing. Also apparently people say to use CLIP 1 on Anything, which this is based on. So go to Settings and set the clip skip in the middle column to 1. (It should be 2 for NAI). Remember that the __pose__ is a wildcard used by Dynamic Prompt extension which I linked in the /qa/ thread.
(huge penis:1.2), large testicles, sketch_lines, nsfw, masterpiece, 4k, 1girl, big belly, loli, huge breasts, (burger:1.25), __pose__, , lisa frank, Coby Whitmore, (patchouli knowledge:1.3), hat, flustered, cowboy_shot, facing viewer, from above, beautiful detailed restaurant, mcdonald's, (colorful)
Negative prompt: twintails, (extra nipple), (extra breast), out of frame, amputee, mutated, mutation, deformed, severed, dismembered, corpse, pubic, real, photograph, poorly drawn, bad anatomy, blur, blurry, lowres, bad hands, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, monochrome
Steps: 26, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 2790887116, Size: 576x768, Model hash: 4aa445b9, Batch size: 5, Batch pos: 1, ENSD: 31337
No.325
>>323On second thought, I tried introducing the hypernetwork into that chunkymix and it actually works really well at reduced intensity. So, now I'm training again from scratch with NAI and it's working well and the preview images do seem a lot better... I think?
However, it's doing a worse job of specifically that Power girl from Chainsaw Man when she pops up since (she's labeled in one of the training images).
However, chainsaws are no longer appearing in the background so I think that's a plus?
No.328
>>327actually I'll post this in the /qa/ thread
No.374
>>326What in the world happened with that sinen_nai 4500 one... she's pulling off her boobs...
No.482
>>481Dang this looks to have turned out very well. I can imagine that the pose training was probably one of the toughest parts.
No.483
>>481>this is what happens when you get moderated on kissuWhat did you get moderated for?
No.484
>>483No no, I mean that is an anthropomorphization of the process. I wasn't moderated at all, but that's what it looks like when someone is. Maybe it's hard to tell since these mixed models don't do specific characters well without secondary training things, but that's Verniy the character doing someone in the butt.
AI is weird, if there's more than one character it's impossible to do individual traits, like hair color. You can add "green hair" and then "purple hair" somewhere else, but it's random if and where the purple hair will show up
No.485
>>481>>484Can't believe this is AI art, wow.
No.491
>>485Compared to the other stuff that's out there the impressive thing about the ones I generate is the accurate portrayal of the penis from behind. The quality and fidelity of people doing normal non-penised women have a huge advantage.
Normally in this AI stuff a penis from behind looks the way that it does from the front, so it gives people penis tails. I tried to will 'penis tails' into becoming a fetish, but it didn't work. It would need some well done doujin first I think. Rarely, it does actually do it correctly, though, but the other parts of the images being good is also random so the odds are not in your favor. Well, the furry model was good at it, but I tried making merges of it so it could somehow use the furry penis knowledge on 2D girls, but it always came at a severe price of completely destroying the faces. (which is a theme it seems)
No.499
>>498How is that impossible in 3D space?
No.500
>>499I dunno, stuff like this looks like the giver has a long, misaligned torso, doesn't it? Or maybe I'm too paranoid about this stuff.
Maybe I should try it on the furry model and see what happens since they have much better genital anatomy by default
No.502
>>500Not that I have any experience with this, but that pose looks very doable to me.
No.504
I reckon I could do that pose.
No.516
>>511Nope, I know exactly know you feel. After dozens of hours of research and work I can generate Kuon which is something I haven't exactly kept secret.
But, something I have kept secret is that I generate lots of images of her that I keep to myself. It's a very special and warm feeling. I was very hesitant to do it at first, but I let go of those inhibitions and I feel at peace. If I can help someone else capture that feeling please let me know how I can help and I'll teach you everything I know.
We are far off from something like the holodeck with tangible environments around us, but I can't help but feel that this AI stuff recently is opening many doors that I did not think would be opened in our lifetimes.
No.517
>>516>We are far off from something like the holodeck with tangible environments around usIf you like reading, we're not too far off from a written version of that. How you can basically have never-ending non-lewd and lewd reading material of exactly what you want is just gonna keep getting better, quickly.
No.518
>>516Also, if you feel that strongly about her, I apologize for making so many crass jokes
No.519
>>516Congratulations on acquiring your waifu.
No.520
NOTE: This isn't an AI image. Read on...
I'm not that far gone yet, I just mean it makes me smile and it's nice to have an on-demand generator for this stuff. People can and should post as many dumb and crude jokes or edits as they want. That's kind of why I was hesitant to say anything.
I finished training a model based on gurimjang's art. I really, really love how he does the shiny skin on his brown girls.
This is what his work looks like, but you may have seen me post it on /ec/ before. The warrior Kuon I posted yesterday used it:
>>>/ec/10617Again I feel conflicted. I'd love to contribute to the community on 4/h/ that have helped me learn this stuff over the past couple months, but I also respect any artist I do this on so my conscience doesn't allow me to do so. I'd share it with a kissu person on request, but I won't just post it for free download.
No.521
>>520Using the exact prompt that I extracted and edited from that image...
It doesn't do weapons very well at all, but you get the point
No.522
>>517From what I've heard the main issue is the idea of permanence or "memory" if you were to humanize it. It won't remember what was said a few lines ago, much less 500 interactions ago. There's also something like a "leak" that builds up over time that eventually breaks them the longer you run it without starting over. Have you run into anything like that?
No.523
>>522It depends on what your using I guess. NovelAI has a "memory bank" so it can remember important things but it also remembers stuff without it for me. I havent encountered the memory leak though
No.526
>>522The one noticeable mistake was that I was a generating a story with asian characters and it described her nipples as
PINK?!?!
No.535
>>534Just use my rule of thumb regarding hobbies, if it stops being fun or creates problems with other things, just stop doing it.
No.536
Yesterday I spent 12 hours combining, comparing and trashing models while taking notes (poorly).
I figured the best way to figure out how this selective merging works was to combine the current hotness (AbyssOrangeMix) with the furry model aptly named Yiffy. Furries really know how to draw penises and detail in general, but I want a hybrid approach since I don't want furry creatures, but humans with penises of Western furry quality. Oh, the horrors I have seen.
Pic related is an example of a subtle change. Seems perfect at first glance, it's got the higher quality penis, but, wait a minute... she's colored like an animal with a belly that's lighter! You might be thinking tanlines, but after doing a bunch of testing this is a pattern that repeatedly shows up. Also that loose skin at the base of the penis is a furry thing for sure. I flew too close to the sun again.
Also trying to figure out where the tag system lies somewhere in this thing. This is tagged with extra emphasis on Elizabeth Bathory from Fate which seems to be the perfect amount of fame for something that works great on danbooru-based models but not so great on others (like furry). It gave her generic horns and the proper hairstyle, but the clothing and hair color isn't there. The horns are really important, too, so I need to find where this stuff is located. She seems like a very good character to test the veracity of the system since I'd expect the furry model to be full of horns.
No.612
>>611even generated pasties
geeeez
No.615
>>516>let me know how I can help and I'll teach you everything I know.I.... This.... My heart.
I'm not sure I'm worthy.
>>534What I'd do is first of all look at the patreon pages of established people or look up their history and see how they started out. In the meantime starting to post stuff on Pixiv shouldn't do any harm. Unless your old account has a history linking it to other stuff, I think it's fine to have just the one, same as plenty other people.
Uncensored stuff, if it's intended to be public you could post it on catbox, mega, or a drive folder, right? If you want it to be explicitly under your pseudonym account then I suppose yeah, mastodon or twitter, I imagine the latter would reach more people but I honestly don't know how things are on mastodon. Could combine a file host with a blog to have the best of both worlds, dodging twitter's compression stuff.
As for commissions and requests, V's gone from patreon to ko-fi so I guess you can ask him. There's a lot of videos about it on youtube, personally I'd only check them out after having looked at established people's history as above since that kind of stuff can be overwhelming.
No.620
I feel like I'm getting close, but now they have more prominent noses. Alas, the increase in Western furry quality along with some bleed from RL stuff has given them noses. By itself it's fine (and I like cute noses), but my Utawarerumono LORA can't overwrite them and that means I need to create one with separate merge options so I can generate accurate Kuon images.
After so many tests I can't seem to eliminate them. As I move the sliders for more furry/RL influence to give them better shading and hands, the noses appear.
I've tried putting "nose" in the negative prompt on generation and it doesn't work.
>>615Yeah, I've decided that I think I'll start organizing stuff now into folders for batch uploads, but since it's pixiv I must also defile it and censor them as well.
No.622
>>621and this is the AOM_nsfw checkpoint which I used as the base. I like it, but it's not as good as following prompts and it looks kind of foggy. The penises don't generate as well as I'd like, either. Meanwhile the furry model can generate them from a bunch of directions perfectly fine which is what I was trying to get, and I think I did succeed there.
The limitation of this model is that if you set the CFG higher than about 6 it begins to break down somehow, and a higher CFG is how closely it follows your prompt. If you just prompt "naked girl" then sure, it'll generate some crazy stuff, but if you want specific poses in specific clothing in specific locations? Nope, not going to happen.
In this example it completely ignored "detached sleeves" while mine at least got it in two of them and drew regular sleeves in another. But why is it better at doing 'orange hair'...
No.623
Oh, and you're probably thinking "hey, these images look a lot better
>>611 " and I fully agree, but unfortunately the slider I found that controls its ability to follow danbooru tags relating to pose and composition also greatly hurts its quality.
The 50 other merges I've made since realizing this have been trying to bandage up that hole to very mixed results.
No.624
I'm trying to understand why specifically OUT04
>>537destroys the ability to control the mouth. The first theory was that it somehow switches to using that e621 furry place's tags, since the furry scrapes are from that, but it doesn't obey their tag ("mouth closed" instead of "closed mouth") either.
For some reason mouths will always be open when I have this slider far enough and you can see it in earlier images in the thread.
also holy crap look at the water here
No.655
>>650now create pixiv account(s) and post art with similar style for each account
No.663
>>655No way I'm going to make multiple accounts, but I think I'll rename my existing account and start posting stuff that matches a theme and uploading it in chunks. As for name, I think I'm going to try and combine the first futa stuff I saw back in the day into a name..
Well, time to think of it now...
>>656That's a big leap from putting in prompts and obsessively creating merges. But, I want to do it with 3D modeling stuff so who knows...
I still need the motivation to do it, though, which is what has kept me from doing anything at all
No.666
>>650Very good pic, just wanted to say that I really like it.
No.671
>>666nice GET
Yeah, autofellatio'ing nature sprites is pretty great.
it's nice when you find a combination of artists, traditional and modern, and a concept that all blends together really well
No.681
>>679Yeah, that's the one.
Been trying to think of my pixiv name, but having trouble. I want to combine two words associated with the first dickgirl images I saw so long ago.
For some reason one of them had "ZAYRE!" in text, and another one had Lilith Aensland from Darkstalkers.
These are the names I'm thinking about:
Aenzay
Zayaen (looks Korean?)
Aenre
Zayrith
Zaymor
I wanted to do "Zayre", but it's apparently a chain of supermarkets that out of business in the early 90s and I don't want people to think of baby senpai nostalgia when they see my name.
I could also try think of a name, but I'm terrible at those. I could try to generate a bunch of text in images and see if I see anything? Dunno....
Why are names so difficult?
No.684
>>681Zaire is also the old name for the Democratic Republic of the Congo.
I think you'd have more options if you further spliced them and played around with their ordering, to get something like Layrn or Dayr. You could take a look at examples of Middle or Old English for that sort of enchanting fairytale feel it's got.
Thought of Zslan too, that one sounds hilarous to me. Looks kinda Polish.
>>682That's what happens when an old term gets refashioned into a quasi-buzzword.
No.685
>>684>Zaire is also the old name for the Democratic Republic of the Congo.Yeah, I ran into that one too. Someone linked this in IRC so I'll try messing with it later:
https://www.spinxo.com/ No.776
Decided to make a new merge with some of the other new merges people have made.
But, I decided I need a good variety of stuff to see at once.
So... behold! This is what I generate for each new merge test. I tried to get a good variety going. I'll also use this, maybe with some tinkering, to run a mega test of artists LORAs later on.
1. Torso focus: body detail
2. Muscular, sweat, seiza. The seiza pose is something specific to booru and other things don't recognize it. It's a good indicator of how the checkpoint is reading tags)
3. Colored skin, dark skin, demon, fat, magic. Some of these checkpoints really can't do dark skin, so it's actually quite important to test it. Colored skin should hopefully give it a non-human tint, but it's kind of unreliable
4. Flan of a certain body type butt. After a bunch of thinking, Flan seems like one of the best characters to use for recognition because the wings are extremely unique. The butt shot also shows whether or not penises are properly generated from behind (this is really hard to get)
5. Face focus. Not much to explain
6. Mermaid has a whole lot of concepts rolled into it including colors and body parts. Furry models do this a lot better as you may imagine
7. Kemono Kitsune. This gauges how anthropomorphic ("furry") the mode is in comparison to the other generations that have furry traits in the negative prompt. This AI gen is starting to warm me up on kemono stuff as long as it's not too furry. Uh oh.
Oh, and I never did the pixiv thing. Bleh.
No.834
>>833also that image shows one of the most horrific-yet-not-too-u
ncommon things: belly buttons on a fat belly replaced by a nipple. I should have censored that before uploading...
No.840
>>671Oh yeah, how do you get the skin to be a nice shade of blue like that? I always end up failing to achieve this
No.850
>>836What settings are you using? Which model (is it mine?)
Is it an artist LORA? Etc, etc. Misplaced/misdrawn nipples aren't too uncommon if you have one or more LORAs loaded and for some reason they aren't interfacing with the model smoothly.
I guess you could upload it to catbox so I can see the metadata, too.
>>840It shouldn't be too hard unless you've got strong character tags (or LORA) working against it.
(blue skin:1.4) or (colored skin:1.3) and stuff like that makes it pretty easy, at least on the merges I've made.
No.852
>>850Ah, I was using the LORA of two artists at once during that time so that's probably it. Although it did fix itself when I did the stitching so maybe I just went down a bad path? Also yeah, the kissu model, it's working extremely well.
I think my issue with the skin was more to do with that I was trying to do it based off img2img, and since the skin wasn't already colored it was extremely resistant to change, when I generate stuff from just a prompt I notice the issue isn't really as present. Although something funny I noticed while figuring out the stitching specific body parts in img2img was also that just selecting the skin and using a hue/saturation filter on it would give me the same desired result as well which is how I made >>841 from the base.
No.855
I have a lot of thoughts about text generation AI. Do you care at all? I mostly use novelAI (not for images dont worry).
I will say, it's telling that it naturally wants to generate smut when you aren't trying to do so
No.856
>>855Not OP, but I'm really interested in text generation AI, especially the offline models that are supposedly extremely powerful. For the most part I just use NovelAI for now too while I wait for kobold/agnai to support 4bit LLaMa models as my GPU isn't really strong enough to do 8bit. It's fair, but my biggest issue is that my aspirations for text-based AI are much higher than 4k tokens on NAI can achieve. I want to write entire worlds of lore and maybe even import them if I can, but just a few entries usually nets me like half the total tokens and it sucks.
No.857
>>856I bet you in a couple of years that will be much more reasonable.
I would recommend using it to augment your skills right now that it is limited to you. There's been massive interest in worldbuilding so there's software specifically for that, perhaps you could save it in chunks?
No.858
Also, it will always be overshadowed by visual AI porn, but the smut potential with text is ridiculous so far. It can do fetishes pretty well and even mix it with the setting, it made a good Nier Automata porn
No.859
>>857In a couple of years I'll have a 50 series so yeah it'll be reasonable
No.1076
>>855I've recently started to mess with it a bit, but I've posted about it on /qa/ already and gave my thoughts here
>>>/qa/106919 (and its follow up post) and here
>>>/qa/107086I've created a few characters/scenarios and image sets to go with them, such as chubby Waggy here. Alas, I still run into the issue I did months ago where too many things seem to be competing in an area of the screen and it's difficult to impossible to get them to generate correctly. Specifically: mermaid bottom, penis, and a belly. I guess it's something you're supposed to use inpainting to solve, but I still vastly prefer the "generate a bunch of images and go do something else and hope you get lucky" method.
>>878Hmm... well with the script someone mentioned on /qa/ a few weeks ago about selecting image prompts in different parts of the image, it might be possible already. I tried it with Furry Patchy and talked about it a bit, but the VRAM requirements are absurd so I don't do it. I guess there's a chance it's better now, though.
No.1144
Been trying to make an improved Kissu Megamix, but it's hard to find improvement. I downloaded some new models to try mixing in, and this one seems like a decent ingredient:
https://civitai.com/models/8281/perfect-worldI decided on merging it with the furry model with this ratio: 1,0.9,0.7,0.5,0.3,1,0,
0,0,0,0,1,1,1,1,0,0,1,
1,1,1,1,0,0,0 and it seems decent enough as that result, but it doesn't have the cute appeal. So I'm taking this one and trying to find the right ratio to mix it with grape to see if I can find a gain somewhere.
I've tried the newer models of the 'grapefruit hentai model', but none of them impress me. I've also tried mixing in the 'Cross Kemono' models, but they sacrifice too much. I'm still not sure if this will bear fruit.
In this set they're in sets of 4 squares and Kissu Megamix is the top 4. Alas...
No.1145
>>1144Ah, to clarify, none of anything in that image is a downloaded model. They're all mixed I've made and it's kind of impossible to describe what they are. I've give them names like [PerfectYay Grape4 134 Mix] which have meaning to me but surely none to anyone else.
No.1193
https://www.pixiv.net/en/users/3727149/artworks?p=1This is a pretty amazing model, but do you think models heavily trained obviously on a specific artist like this are bordering on plagiarism?
No.1198
>>1193I don't know any of those artists, but I heard that pixiv (or fanbox or something related) has started to take action against those that are just blatant copies.
Even though this stuff takes such little effort compared to actually learning to draw, people really aren't using it to its potential at all and are just using one artist at a time at full strength. The true enjoyment comes from combining different artists, anyway, so people are really missing out.
Anyway, I uploaded a chart and I didn't want to edit in this little guide so I'm just posting it here. The images are in rows with each different checkpoint being a horizontal line.
https://mega.nz/file/DKJEkSbJ#k38NVTciBEz7ApitrTHqXJiI6M8dAVlbaucfVSYgZx0(MASSIVE WALL OF DICKGIRLS WARNING)
When I do model merge testing, I just name them with the stuff I've used, so the bottom row is "GrapeEaster MonsterCock" because it has, yes, Grapefruit Hentai Model 4, Easter Furry model and, uhh... "Monster Cock". The monster cock model is pretty interesting because even though it is not labeled very well in regards to characters or poses, by using specific outputs in the merge (OUT3 and OUT4) I was able to get its penis influence without severe model damage.
My pride and joy (kissu megamix), which I am actively improving upon, is row number 8 and then kissu megamix v2 is immediately below it. After my recent testing I don't think v2 is worth keeping around, but it'll stay on the server for the possibility that people like whatever it might do best. You may notice it's the only model that can come close to rendering a penis from behind (last two images in each row) without going through extra steps like img2img editing or other annoyances. Such is the power of furries.
It did hit the face quality a bit and you need a tag like "nose" in the negative prompt, but as a side effect of using the furry model it gives it colorful, flat shading that I really prefer anyway. You'll notice two sets of "real life" prompts around the middle (the blonde in a castle and the beach butt) and the third row from the top is one of those "2.5D models" that are mixture of 2D and real life that is closer to RL. (all of these 2D models have real life data in them, no way around it since it's based on SD).
Do you guys think something like the first and sixth row are worth making a model out of? (the nice painted look)
Models are kind of big in file size, but the main problem with them is that SD can only have one of them loaded at a time. We'll have to test how an online system deals with someone switching models while someone has something queued, because I don't think it'll work well with the base SD UI
No.1199
>>11986 > 2 > 1 > 8 > 4
My top 5.
No.1200
>>1199I could try to make something with the sixth model (ColorBox), but I'm starting to see limitations on it, like how it defaults to a blue skirt every time and all body types are identical despite many different prompts
No.1207
>>1206I don't get why penises are so hard for AI to do right.
No.1209
>>1207The ones from the front aren't too bad, at least in the mixes I've made. I just assume there's very little data comparatively to cute girls, so it's not nearly as good as faces or breasts and other regular parts.
Once you add a LORA for penises it's a lot better, although that usually alters the other visuals a bit, too.
No.1210
LORA for small penises
No.1211
>>1210I have one, yeah, but it doesn't work too well (or maybe it's specific to my merges). The way stuff works is that to make sure a penis appears you add something like (penis:1.2) and that also seems to increase its portion of the image which means it will be bigger.
The alternative is that the penis is only half formed or there's only balls and it just looks awful (and that still happens even with 1.2)
No.1212
>>1211>I have one, yeahcan i touch it
No.1213
>>11982 and 6 look great and are very distinct from the rest, while others like 4/5/8/9/etc I had a much harder time telling apart from each other. I feel like 8 has something going for it but I can't put my finger on it. Parts of 7, particularly the mermaid and Remi, has this sort of grayer lighting that's pretty cool. In general I prefer the ones that go easier on shading, since that aspect can seem very AI-y to me. In the mermaid example I'd take 11 over 9.
It's interesting that in some scenarios a model can look fairly different from the rest and in others be closer to the average.
3 though, kill it with fire.
>>1206In general it makes sense that the ones that stand out are more narrow in range since their training material is more specific, right? Unless I got that wrong.
No.1253
>>1252Damn, the futa cocks are still a pain though I guess. Wonder how some people are able to make them pretty much perfect without errors and even pose them to an extent.
No.1265
File:16.png (39.76 MB,3456x9216)
>>1253These are just 2 images each out of a specific seed. It can generate them pretty well overall. They're not going to be great as-is. People often make use of LORAs for penises, but I won't use them for comparison purposes.
People also generate a lot of images or do the img2img thing for repeatedly generating a specific part of an image until it looks good. You're not seeing any of the process when someone just uploads something on a site.
No.1270
File:25.png (14.14 MB,2304x4608)
25
No.1271
File:A.png (14.11 MB,2304x4608)
A
No.1272
File:B.png (13.97 MB,2304x4608)
B
No.1273
File:C.png (14.47 MB,2304x4608)
C
No.1274
File:D.png (14.15 MB,2304x4608)
D
No.1275
File:E.png (14.34 MB,2304x4608)
E
No.1276
File:G.png (14.88 MB,2304x4608)
G
No.1277
File:G2.png (14.25 MB,2304x4608)
G2
No.1292
>>1291Yeah aside from peach these are for sure really nice looking. No idea why the model fucks up so much with her, maybe too much western art? This leans into 2D-ish more without being too saturated like some of the other models, although I don't think it'd hurt to have one of those really "TV-Anime" style ones in the mix as well.
No.1293
>>1291Actually now that I think about it, probably more like this:
A. General NSFW (with furry capabilities)
B. General Cute (focus on cute faces and recognition)
C. Kemono
D. Watercolor
No.1295
>>1292I do have a theory.
Look at the second image of the third row here:
>>1270But, it is generally going to be bad at the "looking back" face as that was a weakness of the original kissu model, too.
It's probably better now, but I haven't tested it.
But, man, am I happy at how good the penises are. This is expensive furry commission quality.
No.1304
>>1303Oh, boy, you know what else is controlled by these sliders? The natural kemono-ness of the Kissumono Mix.
uugghhhh yeah, I have to rethink this...
I think we're going to have to rely on the kemono hypernetwork or maybe kemono artist loras to do furrification of characters. It's what I did for my Furry Patchy mini-adventure thing test thing on /qa/ so it does work pretty well
No.1386
I normally dont care, but this is too much because he's charging for it while practically cloning the art feels wrong.
https://www.pixiv.net/en/artworks/110136910Its also weird to me that people still care about an h-manga all these years later enough to make fan content of it
No.1389
Do you think the frustration with text generation getting neutered will inspire to people to learn to write good porn again
No.1396
>>1389Hmm, probably not. I think the text stuff will probably create new porn writers, as even I myself who never would have considered myself capable of writing such things have been doing it in small amounts to move the text generation along. The advanced online models (which are increasingly neutered) were great at writing stories along by itself, but as you move to the weaker models the user has to do more and more work.
No.1397
>>1386Its weirdly inspiring in a way that these hardcores are coping with the disappearance of this artist by making their own sequel CGs with AI. Surely the original artist would be more flattered/touched they care that much than just feeling like they stole his art, right?
No.1398
Honestly, its weird how muted artists themselves have been on AI lately
No.1399
>>1398>Honestly, its weird how muted artists themselves have been on AI latelyI'm glad that they've shut up then.
No.1422
Do the image models still freak out and intentionally break if they think you're making loli/shota with it
No.1423
>>1422No I looked it up it doesnt look like it but shota are probably a bitch to tune for as what I see on pixiv they either look like toddlers or lesbians
No.1438
>>1386we're on an IB but you still post URL I'm not gonna click
No.1441
The nice thing about gacha sluts is that there's thousands of them and and hundreds if not thousands of LORAs, but I know nothing about them so they're just blank templates that I can turn into anything. I can't do that with video game or anime characters because I think "she wouldn't do that or wear that".
Three weeks of model merging experiments and I think my final lineup is about 10 models.
I'm going to do some more experimenting, though, as I'd like the appearance of the last on the right with the colors of KISSUMIX 8 UKA95 and the shading of the one first on the left. This is why I keep getting sucked back into making merges...
>>1422Huh? If you're using online stuff then maybe? I don't have experience with that stuff, just local generation
No.1442
>>1441My experience doing that was quite some time ago
No.1502
can you do kuon with two tone hair
No.1510
>>1502not sure if that's a tag, but I can try it later...
No.1512
>>1511oops meant to reply to
>>1502 No.1517
>>1515It does look pretty nice.
Outside of fine-tuning these models, has SD itself progressed recently? Also curious as to how the Kissu UI is going, if I may ask.
No.1519
>>1517Oh and SD itself? Well, the SD company itself made the XL thing which is theoretically better, but in reality it doesn't have the Novel AI leak to build off of so it's very slow going. The VRAM requirements are also higher and just to train a LORA (like if I wanted to add Kuon) I would need like 24GB of VRAM. I doubt anything like the NAI leak will happen again, but we could hope.
As for people creating new plugins and stuff for 1.5? I don't really monitor it, but I look at extensions now and then and there really isn't major that sticks out to me. When you look at the 4chan threads for this stuff nowadays there isn't much information exchanged like when it was new. Presumably people are like me and just moved on since there isn't anything to figure out.
I'll need to look at the animation stuff soon since there's a way to use the previous frame to help influence the next one in an attempt to make it more consistent, but it's still a mess. (since you're making lots of frame at once it also uses a lot more VRAM and it's slower)
No.1521
>>1518>makes me feel like I've gone crazyWell, it's definitely hard to match your enthusiasm. I'm still fairly interested, that's in part why I'm asking, I just don't want to use it publicly. (Also depends on how well the LORA still works.)
>>1519Hmmm, not much then. At least this prompt stuff sounds like it could give some funny results when used to mix facial expressions, if it doesn't produce shoggoths that is.
No.1522
>>1521>Also depends on how well the LORA still works.It should still work about the same, POSSIBLY (maybe) better. I need to go back and retrain stuff probably since the ways to do it have likely improved
No.1540
>>1539ponytails?
or maybe it's eris from mushoku tensei?
No.1543
>>1541then a shab I see posted on kissu but from an anime I haven't watched
that harem one
kanojo mo kanojo?
or that megami terrace cafe one?
No.1544
>>1543Nope, it's a concept and not a character. The bottom-right image in
>>1542 removes all the mystery. This image is the exact same prompt and seed, but without using lora weight. Using specific lora layers allow greater customization and often better quality.
This is something that will carry over to other loras.
Oh yeah, I forgot to mention the reason I'm doing this again is the extension I use for lora weights changed how it's set up so I had to redo this testing. Not sure if it's better or not...
No.1545
>>1544and here is the final result. (spoilered so people can follow the image chain if they want to, it starts at
>>1535)
I'm not going to mess with the numbers. You can put in a number between 0 and 1, but I only do 0 and 1 because I already get annoyed by it.
By using Lora weights you can take concepts which are overpowering and tinker with them so that they can blend in with others. The
gyate face was one such concept as it was clearly trained to show only the head, but by selectively eliminating layers you make it more useful. COOL!
No.1547
>>1545hmmm it becomes more visible in
>>1539 but only with the left violet girl
is it an angle thing? how does it look like with tags like looking_to_the_side or sideways_glance?
No.1548
>>1547>is it an angle thing? I couldn't tell you. This seed stuff is a mystery to me. Just like when you prompt normally, there's a chance it will just be completely ignored or even become really strong.
I doubt the gyate face is possible from the side since the face is specifically drawn at a specific unnatural angle
No.1791
actually I should redo this upload set in order... one sec...
No.1796
i like your taste except for the futa
No.1798
>>1797The fat. If they were a bunch of plump shabbers that'd be perfect.
No.1801
>>1776Without 'pussy'
(also you can see how I turned on the post-processing face option here in comparison to the 'original' I'm replying to
No.1804
>>1803without it... didn't really change much.
Actually I'll experiment again after uploading this
No.1805
nice
No.1806
>>1804oh wait, I just remembered that I have 'pussy' in negatives by default since it naturally conflicts with 'penis'.
Yeah, that was it.
If there's a specific image style you like here I can type it out and throw the LORAs on the kissu merged model IP thing so you can prompt it yourself
No.1880
>If there's a specific image style you like here I can type it out and throw the LORAs on the kissu merged model IP thing so you can prompt it yourselfordered from most liked to least
>>1806>>1809>>1807>>1808 No.1896
File:K8.png (542.09 KB,832x1216)
Manga-styled.
No.1911
>>1896huh that really doesn't look that bad