/bbwdraw/ - BBW Drawn

Deviantart, 3D, scribbles

BBW-Chan is supported by simple text or static image ads from a-ads.com.
Please consider whitelisting us in your adblocker if you enjoy your stay!


Mode: Reply
Name
Subject
Message

Max message length: 9999

Files

Max file size: 10.00 MB

Max files: 6

Captcha
E-mail
Password

(used to delete files and postings)

Misc

READ THE *RULES* BEFORE YOU POST!

[ / / ]

NovelAI Diffusion Thread 3 Anonymous 11/02/2022 (Wed) 03:24:48 Id:370246 No. 122769
Hello, everyone, and welcome to the third thread. Last thread got over 1000 images before it hit the bump limit, and we discovered interesting things about how the AI behaves.
I don't mean to offend anybody here but is there any way that we could at least relegate all AI "art" to a single thread? Just saying.... it sucks.
>>122771 That is indeed the purpose of this thread. Around 1000 AI generated images were posted here rather than anywhere else during the last thread's lifetime. I can't control what others post elsewhere. And frankly, plenty of non-AI art that gets posted on this board sucks even worse. There's still lots of good stuff, and I'll agree that AI art as it is lacks personality at the moment, but it's better than circle-tool and children's DeviantArt monstrosities drawn and posted by the needlessly confident.
>>122769 Quick knowledge recap: https://rentry.org/voldy https://rentry.org/sdupdates2 https://rentry.org/anime_and_titties https://rentry.org/faces-faces-faces My known bbw-favorable western styles ((X art style)) in front of prompt: Pierre-Auguste Renoir, Gaston Bussière, Édouard Manet, Daniel Ridgway Knight, Clara Peeters, Andrei Markin, Michael Garmash, Duane Bryers, Henri-Pierre Danloux, John William Godward, John Collier, John William Waterhouse, Eugene de Blaas, William-Adolphe Bouguereau, Adolph Menzel, Alexandr Averin, Alan Lee, Albert Lynch, Albrecht Anker, Alyssa Monks, Anders Zorn, Andrea Kowch, Andrew Atroshenko, Anne-Louis Girodet de Roussy-Trioson, Anton Mauve, Arthur Hacker, Helene Knoop, Briton Rivière, Dean Cornwell, Rembrandt, Goya My known bbw-favorable eastern styles (many, many more in the anime guide above): kanon, ruu, rossdraws, Nitroplus Fixing eyes, from least drastic to most drastic: "deep pupils", "wide pupils", "bright pupils", "beautiful pupils" Negative prompt for fixing eyes: (lazy eye), ((heterochromia)) My default negative prompt: (lazy eye), ((heterochromia)), lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, bad feet, extra limbs, multiple navels, two navels BBW terms that aren't booru tags that are known to some effect: love handles, belly overhang, flabby belly, soft belly, weight gain, plus size, plus size model, fat, obese, fat face, double chin, fat arms, fat hips, fat thighs... BBW-friendly booru terms: plump, thick thighs, fat ass, navel focus, navel, deep skin, skindentation, undersized clothes, button gap, torn clothes, midriff, ... slob / eating friendly booru terms: food on breasts, food on body, food in mouth, food on face, eating, food, cake, cream, cream on body, hand on own stomach, heavy breathing, blush, ... SSBBW tips: Generate a bbw using txt2img, then take that result and warp it in photoshop/GIMP to increase the size of the subject. It can be a very hasty warp as long as the general shape is good. Take that warped image and feed it into img2img with the same prompt but with a different seed. Play around with the prompt as desired. Alternatively, use a picture or illustration of SSBBW as a base, or make a crude tracing of it and fill in the desired colors, and feed that into img2img. Here's my take at android 21 that I couldn't get quite right, plus some other images. >>122771 Look, I keep all my generated content in this thread. If AI art upsets you there's nothing we can do about it here. And yeah, there's hand-drawn fetish art that's far worse than this both qualitatively and expressively.
>>122728 Frankly, I couldn't disagree more, AI art is the worst. However it's true that all of bbwdraw is garbage regardless and also that you have done well. 👍
Is it possible to Img2Img black and white photos with the colorlessness being preserved?
Found something interesting. I was watching my Studio Ghibli film collection while doing some other work for my other hobby, as one does, and inspiration struck. Giving your image the Ghibli Artstyle makes for rather cute faces. I'll have to experiment with this later. Some seeds make it look a bit more K-ON than classic Ghibli, but more investigation is warranted.
>>123016 More interesting info: The Ghibli style can be mixed with classic artists' styles for more results. However, the Ghibli also likes having higher Steps (which costs Anlas even on opus tier) or it looks somewhat low-quality.
I'll like to request a progression of either makima because chainsaw man been doing numbers or hestia from danmachi cuz shortstack supremacy a big thanks to whoever does this
I don't know what prompts or voodoo magic or wizardry was at work here but I need more of this. Fuckhuge belly spilling over tight pants, big ol boobers, and just as big of a smile. Dunno if it would work, but if you could also give her dark skin and blonde hair I would likely paint my ceiling white.
>>123020 I assume White Lion is supportive(?)
>>123118 Unless she doesn’t notice(?)
>>123118 habits are contagious, as the saying goes though I have a feeling the big cat didn't need any appetite-stimulating drugs to soften up, unlike her partner
>>123178 2/2 with bonus pics. Who wants some yuri? The last pic genned weird characters, but I enjoyed the belly rubbing and S/M nature of the shot enough that I wanted to upload it anyway.
>>123179 I dont think there's that much Yuri ai content out there so please let there be more
>>123180 That's because it seems to be quite difficult, especially with a nice size difference. My bad gens included characters oozing into each other, them being merged into a four-breasted monstrosity, them having swapped hair colors and lengths (as can be seen in the one pic, although they look different enough they might as well be OCs), as well as the standard weird proportions and extraneous nipples. And since I don't like posting without pics, a test sequence of Fate/Stay Night without specifying a character. Looks a bit like a pink-haired Aoi Tohsaka.
>>123170 on the subject of KF, Maybe Lion and Moose got into an eating contest, but forgot to set a victory condition?
>>123089 Thanks pal. It's a shame that the AI doesn't seem to like tanned/brown fatties per >>123285 but I assume it's simple enough to add {dark skin} or something like that to the prompt, right?
>>123281 The red dress one looks like it could have been an img2img of Lexxxi Luxe/Anorei Collins, I'm curious what the input was, if anything
>>123328 not really, unfortunately. I had to paint on the original image and reprocess it a few times to get that result. First attached is the result of Img2Img with an overweighted Dark Skin tag, as you can see there's nearly no change. Second is with running the exact same prompt without a seed image, now we've got darker skin but it's more of a tan. and the final one is the same img2img but I've cranked the 'strength' as high as I can before it just ignores the root image. closer, but we really loose the shape of her. So, most reliable results I've gotten have been from editing an image and feeding that back to the model.
>>123333 Good eye! I see that too. unfortunately it was late, I was high, and I didn't make a note of what image I used. I recall grabbing something random off the BBW board, I believe it was a low-rez mirror selfie. attached is the first generation, you can see she looks significantly different. (interesting aside, I recall the original photo was a white woman, but the first gen made her distinctly black. Kinda flies in the face of my previous point, but interesting regardless.)
>>123210 I wonder what that entails... Staying there? Going about your day as usual while occasionally returning to do pointless contest?
To bring some over from the last thread.
Edit of ananna0315 art (image 1) from other thread.
>>123378 Idol 3/3. Finishes with some side-view and a half-decent belly grab. For those who also want to see a grabbed belly, I suggest the tag "deep skin," as it refers to skin that squishes when grabbed as opposed to "skindentation" which is squishing caused by tight clothes.
>>123356 It means both of them keep eating because neither wants to lose, but neither of them knows when to stop. Stuffing/belly expansion/breast expansion ensues.
Im trying to get it generate nipples but i can’t manage to do it, how do i?
>>123402 Are you using the right generation set? Im not a user, but there's something like a "animecuratedsfw" or an "animefull" in the mix- using the former will only generate "sfw", so even if they're naked, there's no nips or vag.
>>123402 Use the nsfw and nipples tags. As a warning, nsfw is a very strong tag and may dilute other styles you want, as well as cause errors such as extra breasts and nipples in various places. I sometimes find it necessary to diminish it with brackets. Here are some more artstyle tests without a character specified. Sadly, I could not get another belly grab. It seems to be a very rare seed or combination that gets a good squeeze... I've got a couple of Senran Kaguras (note: Yaegashi Nan as an artstyle did not work), a classical art, Pokemon, and one in Kobayashi's Dragon Maid's artstyle. Interestingly enough, it tended to generate pseudo-Kannas with rainbow hair most of the time.
>>123514 after seeing these lined up, another interesting thing comes to mind: whenever I generate seriously exaggerated features (giant tits, huge belly, etc), the composition of the image shifts to a fisheye lens. Of course this only happens with images that have discernable backgrounds. Just a shot in the dark here, but I'm guessing there's some weighting happening due to a number of fisheye style images (or even photos depending on how this dataset was actually created) that focus on these features. Like, think of the number of BBW models you've seen that have taken advantage of a fisheye lens to exaggerate their features. It's interesting to see that replicated here without explicitly calling for it. blah blah shut up and post more images
what is the artist's name? and does it have patreon?
>>123588 First of all, this thread is dedicated for AI generated art. This means that there is nobody who actually drew them, but a computer (with possibly a GTX 1080 or something). The AI often relies on other artwork as a way of training, or inspiration, if you will. Put it simply, nobody but a computer made these, and no, they're not on Patreon.
>>123521 How do you get results like that. >>123522 And how you get the jizz soak?
Could someone please give it some tube feeding (or even better, forced tube feeding) prompts and see how it comes out? I’m thinking about getting NovelAI to make these sorts of images for the force feed thread, but I’m not sure if the AI will be able to add the tube and other elements in a convincing way.
>>123631 The best results (for the most part) come from starting with a target image. I've paired up a few here for reference. I've also been having good luck by starting with a specific "art style" by naming an artist, then describing the scene in plain language (making sure to try different versions of fat/plump/obese with different adverbs). Throw in a few more tags to focus on specific aspects of what you're looking for, then process it a few times. Sometimes I'll feed the results back into the model, other times I'll find a seed that looks good and tweak from there. As for the jizz, soaked/drenched seems to work pretty well, and I've even used 'cum' alongside it if the wet looks more like water/sweat. Of course, sometimes that's what I actually want, but ymmv. But really, it's all just vibes man. reprocess a few times, change tags, take an image to a photo editor and roughly morph it, feed that back in and start over. >>123655 I've only seen one person get any kind of eating to look right, and I've definitely never gotten a hose or any kind of funnel feeding to look right. That being said, if you've got any force-feeding images to share, I could take a crack at feeding that into the model and seeing what we get. I wouldn't hold my breath though.
>>123731 >>123631 Another example. Started with a brilliant image from BetterWithSalt, ran it through the model a few times to find a slightly different art style to lock into, then keeping that seed I (really fucking roughly) edited it in PS to make a before and after image, then ran those through with nearly identical tags and the original seed. That being said, this one tested the limits of what I'm comfortable doing with this stuff. It's one thing to start with either a photo, a 3d render, or a particularly rough sketch (ideally by a non-artist or someone who's just fucking around), but for some reason taking an active artist in the community and doing this felt kinda bad. Like I was stealing from them, if that makes sense. In the future I'm aiming to stay away from active 2d artists in the scene for source material.
Can you give sayori from ddlc a huge belly
(26.66 KB 600x600 Akatsuki.Kirika.600.2111149.jpg)
Can anyone do a fat Kirika from symphogear?
Do you guys think you can try stuff with trinityfate’s art?
>>123731 I am curious about what you can do with someone as BoBerry maybe with all of this.
>>123959 and an even bigger butt. >>123897 yuh >>123931 ok, one more time. Like I said here >>123734 doing this feels kinda not great. Like, I'm pretty sure I could commission TrinityFate to do this, and they deserve the money. That's why I at least do some style mixing so it's not just a pure clone of their work.
>>123731 here's a few i think the AI could have an easier time recognizing i think it might be better to give images that are a tube connected to some sort of mask (that way the AI doesnt need to sync the tube with the mouth) but i don't have any of those saved
>>123972 Are there more parts to that 2nd image?
>>123975 full sequence it goes too quick imo, i wish he would have made the clothes tearing frames on picture 3 their own fully-drawn page source is magnivorewf
>>123977 Thanks, although I think the pace is fine for being 6 parts. Something more realistic probably would've taken triple the amount of pages to do, if not more. I wished he didn't both with the outie, looks weird.
>>123972 That's the thing, it's not the tube-in-mouth vs tube-in-mask that's the problem, its the fact that the model simply doesn't have enough consistent "stuff in mouth" images to sample from. As many force-feeding and funnel feeding images you may have come across, that's a drop in the bucket compared to "generic hot anime lady" art in the sample set. In fact, I would be deeply surprised if a single force feeding image wound up in the training set. So, the model just looses its mind when you try to convince it to output something like that, because it literally doesn't know what it's looking at. Attached are the best things to come out of all this. It obviously understands rope, chains, and restraints, but extreme body shapes tend to freak it out. It's difficult to wrangle the model to output something sane at that level of fatness, but layering on the complexity of a funnel/tube/force feeding is just too much. Or at least too much in my experience. For example, look at the green images. No matter what I did, it simply refused to connect a line to the subjects mouth. compared to the three progression pics, it had a slightly easier time since the pose was normal (the green one is at an extreme angle, making it more difficult for the model to recognize what it's looking at overall.
>>123983 Would you be able to do a BE sequence on the scale of little/nothing to bigger-than-the-rest-of-her-body? Something like the scale that Inuteikoku/Mura works with, except not necessarily to the point that the body disappears. See also Chomikuplus' "Dawn's Ring Fit" sequence (except she starts busty)
Anyone know good ways to stop the AI generating comically large tits? No matter what tags I put into the avoid section or use in the prompt it likes to occasionally interpret being fat as having massive oversized tits.
>>124011 A good example of the scale for Mura's BE stuff is 318771 on nhentai Dawn RFA is here, encoded b/c it's faster than checking if encoding is necessary or not: aHR0cHM6Ly9lLWhlbnRhaS5vcmcvZy8yMzI4NjE5LzRlYWYyZDViMTcv If anything, I'm hoping it's mutually beneficial- you get practice keeping the AI to some kind of consistency between drastic size differences and I get my primary fetish catered to. If that fails to be the case, though, you don't gotta spend too much effort on it.
>>124021 It's an issue inherent to the model, it associates more things with big boobs than you would expect. Best course of action is negatively weighting tags that describe big boobs and add a "small breasts" to your positives
This was harder to generate than it needed to be. Rem kinda has a double belly button but it looks more like her belly is bent instead. I could not for the life of me get even a decent fat Emilia. Even she is a bit inaccurate but whatever.
>>124011 How are you getting the bottles/held items to look so good? In my experience it tends to generate an amorphous object unless it's a very simple shape to begin with
>>123965 >trini makes sense, i can respect that. though, it is true that the ai just doesn't hit the same as the original artist
>>124024 Mutual beneficiality aside, until I can get this running on my graphics card (compiling the AMD version keeps failing for me for some reason) I can't spend too much time on this. No offense, it's just not my cup of tea and I don't want to burn too many credits. That being said... hoo boy the model fuckin' HATES egregious body part sizes. I decided to do a re-run of the breast expansion I did here >>123514 to add a few more steps. Of note, I didn't edit the nipples on this at all, for some reason I still get a lot of gold or black dot nipples. I overcame that in the first set by manually recoloring them and reprocessing to keep the colors somewhat consistent, here I skipped that set. As you can see here, you can still get some really giant tits with little problem.
>>124134 but then this happens when you flip to room filling. I've never been able to get the model to create a truly enormous body part without photo references, even the previous version still relied on a progressive series of 'shops to get to that size. Taking the Dawn RFA example as a reference, the model either spits out garbage like in the first image, or just blatantly copies the reference when you scale it down enough to get something recognizable. I might do some more tests on hyperfats (again, all apologies, but that's more related to my personal interests), and if I figure anything useable out I'll come back to room filling tits.
>>124056 Let me be upfront: I can't claim to truly understand how this works. to borrow a metaphor from my post here >>122572 we're all just playing around with a machine that we don't understand how it works, banging on the sides of it to make it do something interesting. So, now that I've set your expectations appropriately, my best guess is context and a bit of luck. In the images you're referencing, the first three started with a reference image that had the subject holding a bottle. The last three were generated whole cloth, on the other hand. Attached here are a couple experiments. I've included the prompt at the bottom of this post if you want to play around with it yourself. The only difference between the two sets are the art style referenced at the start of the prompt, and then the resulting image used for IMG2IMG. The first three start with Svetlin Velinov as the art style guide, and I must admit the model tend to be all over the place when you reference his work. He's a very talented artist most known for his gorgeous work on MTG cards, and sometimes referencing him gets rich skin tones and excellent shading on an otherwise 'anime' style. In this case not so much; since a large portion of his work is high fantasy and creature design, I think that when your prompts are mundane the model doesn't have a lot to reference. So, even though these are pretty good, the bottles are somewhat nondescript. Helene Knoop on the other hand is a figure painter whose style is deeply rooted in classical/grecian masterworks. while I can't prove she's ever painted a bar scene, her work would fit right in with 1700's-1800's rococo and neoclassical masterworks and I guarantee there's a ton of pieces that are set in bars and pubs. So, I think there's more stuff there for the model to free-associate with and thus creates more natural looking bottles and bar scenery. Nuts and bolts wise, I go fishing by creating a prompt and letting the model create random seeds a handful of times until I stumble onto a scene that's close to what I'm looking for. I'll lock in the seed and then play around with the settings a bit (and normally add/remove/reweight tags, but I kept them static for this test). Then I'll take the best result, feed it back into the model with IMG2IMG and mess with the noise and strength values. I've generally had the most success with a slightly higher noise setting at this point, as it allows the model to create interesting/more 'realistic' backgrounds, and in this case tends to give the bottles a bit more detail. You can actually see this happen in 'real time' in the three image progression here >>124011 as the first image was my starting image, which I edited in PS and fed back in for the subsequent images. The first image has no background, the second had a high noise setting so it created something, and the third reinforced the second's results. And finally, real language prompting. Unlike most other AI image generators, everyone I've seen talking about this one stressed the importance of just using tags, and trying to hew close to the danbooru tag set. While that definitely gets reliable results, I've personally found having a plain language descriptor as one of your tags tends to create more interesting and internally consistent images. I have no idea why this works, but I encourage everyone to try it. Doesn't have to be flowery prose, just simple descriptions seem to have a net positive effect. The tags used for these: {{NAME OF ARTIST art style}}, a very plump woman stands at a bar holding a bottle, drunk, happy, flabby belly Seed for the first Svetlin Velinov image: 3542259879 Seed for the first Helene Knoop image: 1204829575
>>124134 >>124138 Ah damn, I didn't realize there was expenditure of limited resource involved, my bad. Still, I appreciate that you tried at all, that first set is still very nice, even with the weird beads (maybe it thinks they're piercings or something?). But yeah, you do you, the stuff you've been making already is great. The weight gain sequences and hyperfats still have growing/big boobs so I'm eating well either way
>>122780 Hey guys so I just downloaded the Voldy leaked AI about a day ago and spent all day playing with it. I can get some nice bodies especially after finding some of the hyper and balloon diffusion models but no matter what I do I keep getting fucked up faces where its either blank or minimal detail in the face. How do you guy get your models to have some lovely faces? Do you guys think I'm missing some kind file that helps with making faces look good or what? Heres the fruits of my labor today, last pic probably the best thing I've made on it so far.
>>123982 Hm well that makes sense. I guess not even AI can save us from the nicheness of these fetishes yet. However, someone could probably make a lot of good art by giving it something like "fat girl tied up" and then editing in a tube on the convincing ones.
oh shit. NovelAI dropped a new furry model, and it looks like it has promise for easier fetish stuff. First pair of images: used the same prompt as this >>124146 post, but added the word human at the end. Second is one single img2img reprocess with monochromatic negatively weighted. Third image prompt: a very plump woman grabs her belly, human, ssbbw, flabby belly What's remarkable is that's the first image that spat out. I can chalk some up to luck, but I've never gotten something that fat (let alone grabbing her belly!) on the first try without a seed image. Fourth Image prompt: {{rossdraws art style}}, a obese woman puts her belly on the table, side view, human, flabby belly, large breasts, blue eyes, deep navel Again remarkable thing is getting a belly on the table first try. I normally have to fight the model HARD to get anything like that without seed images. Fifth image prompt: a obese woman with a large belly that hangs to her knees standing in a bathtub, human, {{flabby belly}}, giant tits, brown hair, navel spread, side view, huge ass, {{drenched}} biggest note here is the completely reasonable looking nipples, and while the navel is slight, it's in an acceptable spot. Again, first try. A few key takeaways so far: The furry model reacts oddly to {{ARTIST NAME art style}} tags. like, it clearly knows Helene Knoop, but isn't good at drawing faces with her style (in my few attempts so far, anyway). other reliable art styles I've used in the anime model just seem to throw wild results (Mr. Velinov definitely doesn't draw like a horny Genndy Tartakovsky, for instance). Faces tend to be a mixed bag. Sometimes they're great, other times it feels like last year's level of quality (not good). But overall, I'm getting pretty good bbw/ssbbw results on the first or second generation nearly every time (as long as I include a human tag, anyway). Thank you furry community, I guess, for including so much fetish content in your art.
>>124134 Can someone show me how to make A.I I want to make stuff like this
>>124846 Did you try being born with talent? or perhaps simply learning to draw? Then you could aquire skills over time get better
>>124853 Wrong thread, fuck off
>>124846 correct myself I meant show me how to make A.I art
>>124721 >>124723 High quality stuff, much appreciated! Yeah, unfortunately the more niche a fetish is, the higher chance of being comorbid with furry shit.
>>123170 >>123020 Speaking of them, could anyone try an Img2Img to generate a larger form on their official artwork? Wondering how it will do this kind of render.
(6.63 MB 768x1152 gif.gif)
>>125090 Nice work on the gif bro
can someone made some Hinanawi Tenshi from touhou or more Junko from ZombieLand Saga? thx
Is there a way to run balloon diffusion with other diffusion models on google collabs?
>>125485 also, having better luck with it understanding 3d space. Didn't even have to try to get that 'stuck in a doorframe' look.
Is there a way to download the new furry model locally? Or do I have to subscribe?
>>125570 HAHAha! How about why don't we instead call it what it is? 🤭
>>125572 Look man, I get it. AI 'art' is a blight upon the world and spits in the face of artists everywhere, and I dunno, probably something about degeneracy or whatever? Cool, got it. No one here thinks that this is high art. No one here thinks that these tools replace actual artists, or that any of this is 'good' or 'of value. These AI powered image generators are *at best* a tool, but really not much more than a fancy toy. You're absolutely correct, anything generated and posted here is 'soulless.' It's fascinating to watch these tools develop and it's remarkable at how fast they've improved in such a short time span. But it'll be a long time (and I argue likely never) before any AI can replicate artistic intention. Look at the examples here >>123734 and here >>123965 Both of them start with art made by actual artists, and the results (while aesthetically pleasing to some) don't really say anything. But when you look at the source images, there's a story being told. That right there is the 'ineffable' nature of real art: the act of conveying an idea beyond the elements of the composition. (god it feels so fucking dumb having this conversation in relation to literal fap material, but here we are). Nothing I've posted does that, and I would never argue that it does. So, no. I don't think I'm an artist. I haven't acheived anything of value or substance here. A generous interpretation of what I (and others posting in this dedicated thread to ai generated images, and not any of the other threads on this board, a thread you can easily mute and never have to see) am doing would be similar to that one guy on the quad of your college who's pretty good at devil sticks or the diabolo: an interesting feat that means nothing, but is kinda fun to look at if you like that kind of thing. And much like that guy on the quad, you can easily ignore this. And since I don't like posting without images, here's a bunch of shit renders you can use to prove how awful AI images are.
>>125647 Absolutely, your examples right there can prove what an AI is capable and incapable of doing. double arms, tri-breasts, incorrect anatomy, double belly buttons, I just don't see it as replacing real artists in any way (Though some members of DA are upset at DA's implementation)
Im I bit worried it will satured the medium and well, die down from there.
>>125691 A reasonable concern, certainly. My main counter to that is we're in the 'new toy' phase. Since access to these tools are somewhat democratized, lots of people are playing with their new 'thing' and seeing what it does. Imagine a class of 100 kids that have just been given yoyos. (if you're of a certain age, I bet I just unlocked some memories). Some guy stood in front of them and showed off a bunch of wild tricks, and taught them all how to do a simple trick. For the next month, parents and teachers are sick to death of yoyos, since nearly all 100 kids are constantly playing with them, trying to do tricks and whatnot. But within that first month, some kids stop playing with them; tricks are too hard, they can't get the hang of the basics, it hurts their hands, whatever. in the next month, only 15 kids are still at it, congregating with each other, showing off their new rough skills. Within another month, only one or two of those kids will keep at it, as they either had a natural knack for it, or found the process of learning tricks enjoyable. By the end of the semester, no one thinks about yoyos any more except for Jimmy who's uploading yoyo tutorials on youtube as a hobby. Right now, we're in the first week of the yoyo craze. Teachers and parents represent platforms like DA and communities like this one, all trying to figure out what to do about these toys. in the right setting, they're nothing more than amusements, but can be disruptive during class (posting AI images in threads not for that, shitting up insta feeds) or can hurt other kids (people passing off generated images as art they 'made', generating art in the style of an artist). Over time, rules and processes are drafted and put in place. But overall interest will wane, reducing the need for stringent control. It's an inelegant metaphor, one that explicitly side-steps the moral implications of these AI generators and their datadets because I am not equipped to have those arguments. But these tools exist, and people will play around with them, but many will grow bored. Niche hobbyists will keep dicking around with this stuff, but I promise the ubiquity of shitty generated 'art' will slowly fade. >>125648 I've been dormant on DA for the last month or so, what implementation are you referring to?
>>125444 The response to interpreting 3D Renders has made me really fascinated and I think it highlights one of AI Artworks’s niche appeals that is peering into an alternate reality. A what if rather than something to replace.
>>125727 And by doing that I don’t want to throw our reality away.
I understand this stuff gets a lot of hate, but this is comically bad art. Part of the reason I don't judge this art so much is because I see it comparable to the rest of the art that has always been posted here prior yo AI, but the reality is that I have read anon complain on more than one occasion about how it had been YEARS since he saw an art piece that was worth downloading be uploaded to one of these threads. So I know I am not alone when I say that literally none of this art on this site is good, a handful of it being mediocre at best. The fact that I stay mostly silent on the issue does not mean that the art I have seen here is the least bit to my liking. That is all I wanted to say. I understand the masses clearly enjoy a different aquired taste. I have no problem with them enjoying what they enjoy and to a certain extent I even welcome it, but I am also not going to pretend that it's the best when it's literal garbage; sometimes 40 or 50 year old styles garbage. I have no qualms about it as long as we can agree to call crap crap when we see it without the need to lie, decieve, or manipulate one another. I apologize and I will not comment anymore on bbwart threads. I usually stay in tits anyways and occasionally comment on booty threads.
>>126151 Have you tried any character sheets yet?
>>126152 I LOOK AT FAT ART TO CUM EVERYONE HERE IS JUST TRYING TO CUM EFFICIENTLY REAL OBESE WOMEN ARE USUALLY UGLY I JUST WANT TO TRICK MY LIZARD BRAIN INTO THINKING IM CUMMING INTO A WOMAN OF UNREASONABLE PROPORTIONS MOST FAT ART SUCKS BUT THIS IS GOOD ENOUGH
(296.11 KB 512x512 ex_01.png)
(470.93 KB 768x512 ex_00.png)
(341.63 KB 512x512 ex_03.png)
(506.91 KB 768x512 ex_02.png)
Here are some results of my experimental obese/belly model trained in DreamBooth. I used the modified NAI from China as a base. I'm on a quest to create glorious fatties with AI.
>>126152 Save this for the man baby thread you driveling retard.
something i've been kind of wondering, specifically for NovelAI, if you wanted a specific character, would it be better to enter their features manually, or can you just straight up put their name down?
>>126730 You can name them and if they come up in the recommended prompts then the ai will be able to make said character depending how strong the recommended prompt is
If anyone is willing to part the knowledge unto me what prompts do I have to punch in the actual novelai to get girls in this thread and what's the best method for making progressions cuz I want to fulfill some people's requests.
Should I install and learn a distro to take full advantage of my RX 5700's ROCm cores and get full utility out NovelAI, or should I spread my cheeks for a RTX card?
>>126782 I would direct you here >>122780 For progressions, either use the same seed and keep adding (((emphasis))) around a fattening term, or generate an image, put it in img2img, add more fattening terms / more emphasis. For really big, you'll need to warp the image in a photo editor and then feed it to img2img.
>>126810 What terms would you recommend I really want to get this stuff down like a charm
Some of my attempts from mucking around on the NovelAI website using various methods. It would be freakin amazing to have an easy to understand guide at how to make this shit work better more consistently. I'm willing to pay for the service, but if it's going to be as hard for me to understand as it is now I'll probs axe my subscription. That saying I'm a noob and too lazy to scroll through all these threads about it
>>126865 My brother in Christ, the best info you're gonna find is in the post linked >>122780 there, and in this one thread. It's one thread, you can read the whole thing in like thirty minutes. Faster if you skim through the requests and single sentence posts. Right now, you get better fats with the furry model +human tag, but it tends to be more unstable. If you've got the credits to burn ($10 gets you 10,000), go fishing with a handful of fat descriptors and generate a few. Find one you like? either save the seed and play around with adding {{emphasis}} (each pair of {} brackets makes that tag 'more important' to the model) to a specific tag, like {{flabby belly}}, or save the image and feed it back into the model with Upload Image. Best/more reliable results come from taking an image, roughly editing it in a image editing program and feeding it back into the model. There are no hard rules with this stuff. Experiment and find out what works. We've all shared our findings here, report back with any new discoveries.
>>122769 maybe a mod should sticky the ai thread? that way it's all contained to one thread forever? but also that would allow people to find all information about it in one place? thread 1 and thread 2 already are lost to time now it seems?
bbwchan should get a new board for AI generated shit, none of this is hand drawn or made.
>>127103 That's... That's what this whole thread is about. Do you NOT know how to read?
>>126318 looking great so far. would love to check out the finished model ive been trying to make fats with textual inversion, but mine look horrible. i want to try dreambooth but i have no idea how to get started
>>127105 maybe you should be the one learn to read you twat, im saying get a whole fucking separate section on this site for this shit. this is drawn, this shit is not drawn, so it doesnt belong here. theres enough of this cancer on this site already, it needs to go into its own seperate dumpster like the inflation and pregnancy boards.
Some thanksgiving Iowa
Some Thanksgiving Iowa
>>127249 >>127250 And it was posted twice, cool.
>>127237 It was a mod who made the first AI thread. It was pinned.
>>127292 The main purpose of a thread like this is keeping it localized there.
>>127294 I agree that it is drawn, but feel that AI art really does not require enough threads to warrant a whole category at the same time. The most I can see is a request thread, a help thread, and ones for different AIs, but I feel like one thread or two like the help one we had already does all of that. If you want to see why in my opinion a new category would be a waste, look at the state the writing category is in. I still remember when very active, friendly, and broadly helpful writing advice thread that was here got sent there and almost immediately died.
>>127300 I meant I agree that it is not drawn. I said the opposite by mistake making it look like I disagreed.
>>127300 agreed. there is barely enough demand for a second thread, let alone a whole other board. and people like it, plus its not any less deserving of a spot here than the deviantart circle tool OCs that get posted
>>127237 Both Inflation and Furridom even beyond BBW are as categories broad, with both there are a ton different venues to take from regardless of your opinion on them. AI Art meanwhile, is that simple. I don’t know what’s the point of making a thread dedicated to one fandom when the interest is as simple as the fact it is AI Art. And AI Art I feel like is only really guided by us rather than made by anyone. Better With Salt has a thread because there is a community to be fostered there. There is only one Salt. Novel AI can be anyone.
>>127305 Meaning community is of the AI, keeping it isolated.
>>127307 That is not to say people can’t foster communities or stakes of their own through it, but that these communities I feel will end up with more in common than separate. That overlap in my opinion is a unique aspect of the novelty that is AI Art. That being one that sets it apart.
I’m wondering how Novel AI would apply a softer face and tum to this one. Maybe arms.
>>127328 For the attempt thank you. I was hoping the angle would have lended to straight forwardness. Turns out that was not the case, I agree.
>>127300 You're literally in a thread that's being filled with content. The last one locked at 1000. What the fuck are you talking about. I think the ranting thread is up again, go check that out.
(252.79 KB 503x656 IMG_4644.png)
(295.93 KB 515x689 IMG_4646.png)
(581.56 KB 512x768 IMG_4647.png)
Recently bought NovelAI last night and started using it for quite some time, its pretty fun and easy to use. Here are some AI pics that I made today.
still very difficult to coax good fats from the model. i am using stablediffusion and ive noticed the following gives better results: img2img on fat artwork or crude edits a small descriptive sentence like "very fat woman standing with big belly" along with the rest of your tags avoid using "obese", "huge belly", or "gigantic belly"- they make the bellies very round and spherical, without any rolls or details avoid using "bbw" and "ssbbw" - they dont really do anything "belly focus" is good sometimes but may create weird lumpy bellies avoid "skindentation"- tends to merge the breasts and belly together. "navel" and "belly button" on the negative tags sometimes produces better bellies. play with it "flabby belly" is a godsend and should always be emphasized at least once. "fat rolls", "belly rolls", and "big belly" are also pretty good "plump" is good as well, especially with emphasis, but sometimes invokes a shitty art style or will actually make bellies smaller. play with it merging the hyper ass model with a normal model of course lets you generate huge asses, but also makes girls generally wider and makes bellies larger
>>127443 and of course wider image resolutions make wider girls
>>127341 Content that is fine being here? Imagine if there was an entire category instead of one source thread.
>>127460 Or two, if you count the archival thread lost art.
(507.13 KB 512x768 tmpttil6xcg.png)
(534.13 KB 512x768 tmp2uae9of3.png)
(489.40 KB 768x448 tmpcu08aofm.png)
(2.03 MB 1120x1920 tmp2j4vn448.png)
(4.71 MB 1792x3072 tmpgfcag1bg.png)
Dropping the good stuff I got since when the thing first hyped
Trained a model based mainly on my favorite artists (you might be able to recognize the creators). No img2img necessary. It's pretty accurate to the point that this definitely seems like some kind of violation, so I'll be keeping this one for myself I think. But I thought you should know it is possible.
>>127771 That's interesting. Can you tell me what this is supposed to be for? Is it supposed to look dead and lifeless? Can you tell it to make a hentai scene?
>>127772 you sound like youre about to quit
>>127773 Okay, but what does what you said mean?
>>127776 That you should kill yourself, dumbass.
>>127784 Okay? Why should I do that, idiot?
(369.71 KB 512x512 12313.png)
(313.45 KB 512x512 3233.png)
(441.95 KB 512x512 1231.png)
(369.80 KB 512x512 32.png)
>>127771 JESUS CHRIST this is good... right when i gave up on training my embed cause it spits out picrel id kill to get this or at the very least train one myself is this dreambooth or text inversion or hypernetwork? what are your params and sample size?
I’m hoping Novel AI will some day allow generations in exchange for watching ads or something. If more people had access to it, I feel like there wouldn’t be as much fear surrounding it.
>>127805 doubt it will ever happen. the fraction of a cent they make off of the user watching an ad would not come close to covering the costs of generating an image but whats your gpu & operating system? if you have a modern-ish computer you could just run the NAI model offline
>>127798 This was done by merging the model I made with the Chinese NovelAi clone, Anything-V3.0 at a 50:50 ratio. I used the Dreambooth extension that can be added to the webui. There are tutorials online on how to get it working correctly and how to use it. You may have to update Visual Studio if you're using Windows so Stable Diffusion can import pytorch and torchvision. This model was trained with 470 images with a roughly consistent style for 47000 steps, 100 steps per image seems to work well, anecdotally. Also about 2000 class images. Class images don't need to be a consistent style, just digital art. Went to bed while it cooked my 3090 for about 8 hours and that was it. Without merging the models it gives rougher images but decent body shapes. So the Anything model kinda works like a cleanup or a big anime hypernetwork. Specific parameters don't matter as much since the model will bias toward fat shapes anyway, but here's the prompts for the Miku pic: Euler a, 30 steps, CFG Scale: 11 1girl, ((hatsune miku)), vocaloid, ((blue hair)), long hair, thighhighs, pantyhose, smile, white shirt, tank top, sleeveless, standing, fat rolls, belly, (((obese))), (((fat))), nsfw, thick thighs, round face, fat rolls, fat arms, midriff, big belly, tight clothes, gold bracelets, beach background, beach, sand, ocean, palm trees, tree, clouds, sky, food, holding food Negative prompts are just the usual low quality, bad anatomy, etc I've had some luck training embeddings for specific characters, and it seems plausible you could get similar results to this with a big enough hypernetwork, but the Dreambooth method seems to be the best for shape and style training from what I've seen and read. Good luck
(141.95 KB 1024x1536 styletest1.png)
(4.25 MB 384x576 interpolated (1).mp4)
https://note.com/4m4m1_/n/n8bb0eb792490 https://www.deviantart.com/zachsano/art/AI-belly-art-keywords-Belly-size-chart-939170401
https://www.reddit.com/r/FatHentaiAI/ New subreddit for this kind of art. I think you gentlemen will enjoy. https://www.reddit.com/r/FatHentaiAI/
(454.94 KB 448x704 00083-2561312656.png)
(339.43 KB 448x704 00053-694723940.png)
(421.90 KB 448x704 00028-4000857419.png)
(430.45 KB 448x704 00375-582248387.png)
>>127771 >I'll be keeping this one for myself while I don't agree with keeping models/prompts/tips to yourself, I can't really make you share it since you trained it yourself. It's a bummer because there haven't been any models/embeddings shared for fat related art (that I'm aware of), so the rest of us are stuck twisting existing ones into pretzels and praying to RNGesus that we don't get flesh piles. If you aren't interested in sharing that model, then would you be willing to train one more generically on different sizes (like chubby through obese blobs/immobile)? I have thousands of pics tagged in my local hydrus but since I'm a gpulet, I can't train anything myself
>>123387 cheers >>123210 moose's hair antlers are proving to be difficult lol
>>127810 Nice work. I also have been training some models on Dreambooth based on my favorite artists. <br> <br> I have two questions about your workflow if possible.<br> 1) Did you train the text encoder during your long training session? <br> 2) Did you handpick your class images from the web or generate them using the China NAI base model for your model using a class like &apos;digital art or obese&apos;?<br> <br> Some of my workflow for people interested:<br> I used Shivam&apos;s Dreambooth with a GTX 1080Ti on Linux. Nerdy Rodent has a nice guide on that.<br> <br> I&apos;ve been using a specific class like &apos;belly&apos;, generated images with the Chinese NAI for class data (800 or so), and then gathered training data that has a focus on bellies, specifically works of my favorite artists. However, I had to edit the training data such as removing backgrounds or any odd details becaues Dreambooth likes to pick up these things in the finalized model. My results in shown above are 7-8 images with about 800 steps (images * 100 steps). I also trained the text encoder.
Nice work. I also have been training some models on Dreambooth based on my favorite artists. I have two questions about your workflow if possible. 1) Did you train the text encoder during your long training session? 2) Did you handpick your class images from the web or generate them using the China NAI base model for your model using a class like (digital art or obese)? Some of my workflow for people interested: I used Shivam's Dreambooth with a GTX 1080Ti on Linux. Nerdy Rodent has a nice guide on that. I've been using a specific class like 'belly', generated images with the Chinese NAI for class data (800 or so), and then gathered training data that has a focus on bellies, specifically works of my favorite artists. However, I had to edit the training data such as removing backgrounds or any odd details becaues Dreambooth likes to pick up these things in the finalized model. My results in shown above are 7-8 images with about 800 steps (images * 100 steps). I also trained the text encoder.
>>128012 Those are some decent looking results. I'd be interested to see more like this. No, I didn't do any additional training of the text encoder. It's able to generate the images using only the standard prompts like "belly", "obese" and "midriff", etc. Training the encoder does seem like it would be very helpful if you wanted to generate "blob" or "immobile" type fats as others seem interested in since the encoder has no knowledge of these terms. That's never really been my taste, but that's how I'd go about it. For the class images, I just used 2000 images from my local folder of fat art in a bunch of different styles (not handpicked per se, but it's still art I like) and gave it the class prompt "digital art of a fat girl". If my understanding is correct, the exact subject of class images shouldn't matter as much as the type of image, ie, "digital art", and "girl", as opposed to "photograph" or "man". But it may help a little that it was fat art, I'll admit I'm not exactly sure. Either way, there should be no problem generating the class images yourself. This is the second time I've tried training a model like this and I believe the main difference was the quality and consistency of the training images. The first time I used art from many more artists that was in differing styles, and it couldn't produce the same quality output. Limiting the training to 3 or 4 artists that I really like seems to be what made the difference. Since the images have to be cropped to 512x512, getting a good mix of face, belly and thighs in the training set seems to have helped too.
(464.29 KB 512x768 0.png)
(429.24 KB 512x768 1.png)
(465.64 KB 512x768 2.png)
(420.13 KB 512x768 3.png)
Hot Spring Fox Spirit 1/2
(467.72 KB 512x768 4.png)
(428.35 KB 512x768 5.png)
(453.34 KB 512x768 6.png)
(445.57 KB 512x768 7.png)
>>128024 Hot Spring Fox Spirit 2/2
(118.41 KB 1024x1024 wanda.jpg)
(118.79 KB 1024x1024 cw.jpg)
(216.44 KB 1536x1536 harley.jpg)
(129.13 KB 1024x1024 carol.jpg)
(83.99 KB 1024x1024 ww.jpg)
(115.76 KB 1024x1024 sg.jpg)
I saw these floating around.
Here are some of my tries. Is there any advice on the tags to avoid blurry nipples?
i trained two models in dreambooth- one locally for 1.4k steps over 5 hours (cpu-only mode is slow as fuck but i dont have the vram to put it on my gpu), and another in 6k steps over one hour using the dreambooth colab. the first four images come from the local training. the last three come from the colab training i assume the disparity comes from how i accidentally overtrained the colab model and had to merge it to get anything usable. the local model didnt need merging (but does need luck- i got these images with a lot of luck and inpainting) ill play with both some more and see if i can get something better also >>128024 >>128025 fantastic post
>>127985 In my opinion, someone on Twitter, @Santaroumikan, managed an impressive Ezo, although it seems like it hasn’t reached a stage of being consistent.
>>128102 It turns out their Pixiv has this resource documenting their experiences. https://www.pixiv.net/novel/show.php?id=18702469
(2.71 MB 384x576 interpolated (3).mp4)
(1.26 MB 384x576 interpolated (2).mp4)
(1.60 MB 384x576 interpolated.mp4)
I like making these...
(294.29 KB 512x512 ex_01.png)
(302.95 KB 512x512 ex_00.png)
>>128017 Thank you for sharing your wisdom. Here's some more fatties from my model.
>>128113 That first one looks like you used one of Kurocaze's artworks as the reference.
>>128106 This is fucking dope
>>128170 >03362-646818324-masterpiece, best quality, belly, (bbw_belly_1), (sweat_0.8), school uniform.png) What word do you use ? And do you use "Undesired Content" ?
>>128170 Trained on, or referencing miraclerun. The fifth picture in particular looks traced.
>>128185 >>128195 yeah, its a model im trying to train, primarily on art from miramiraclerun, primesui, vossdpy, and sweetdreamcoffee. i will share it when i get it right, because atm it knows a single pose and can only do schoolgirls
(1.49 MB 1532x640 Marissa 1.png)
(1.10 MB 1533x640 Marissa 2.png)
(378.33 KB 512x640 Marissa 4.png)
(1.15 MB 1532x640 Marissa 3.png)
It's me again, best Touhou enjoyer. Got caught up with personal life stuff for a bit, but I got most of it sorted out and had the chance to just set up and let sd-webui run again. Been reading through the new developments slowly and applying the things that I have liked the outputs of. Got a few Ran's to share first
>>128415 >>128416 >>128417 based posts also theres a hololive prompt list? can you drop a link? id kill for a marine & kiara
(157.19 KB 694x766 qqsdcq.jpg)
(227.02 KB 1116x713 153.jpg)
>>128418 https://rentry.org/3y56t this was the one I used, but there is also this too https://rentry.org/q8x5y
>>128421 thanks, i appreciate it. i'll play around with it
(523.03 KB 832x512 tmp3mrwhyf_.png)
(494.58 KB 832x512 tmprrajm_5w.png)
(530.36 KB 832x512 tmpihadwik6.png)
(681.47 KB 640x832 tmpuika53pd.png)
(672.22 KB 640x896 tmpg0rt807j.png)
(518.12 KB 640x640 tmpbu3bmuj_.png)
>>128601 whoops posted an image twice
(534.99 KB 512x768 1.png)
(516.37 KB 512x768 2.png)
(516.46 KB 512x768 3.png)
(511.98 KB 512x768 4.png)
Goddess Can't Get Enough 1/2 If you want the text that goes with the images go to: https://www.reddit.com/r/FatHentaiAI/comments/zc496s/goddess_cant_get_enough_wg/
(507.76 KB 512x768 6.png)
(473.80 KB 512x768 7.png)
(513.03 KB 512x768 5.png)
(678.16 KB 512x768 8.png)
>>128636 Goddess Can't Get Enough 2/2 If you want the text that goes with the images go to: https://www.reddit.com/r/FatHentaiAI/comments/zc496s/goddess_cant_get_enough_wg/
(289.23 KB 602x549 screen1.png)
(444.90 KB 604x776 screen2.png)
(471.88 KB 604x776 screen3.png)
(461.74 KB 606x778 screen4.png)
(582.28 KB 607x779 screen5.png)
(493.40 KB 604x777 screen6.png)
Ok, so I did a little testing with some other models I found, and the best one to compare to NovelAI turned to be Anything Diffusion. Here you can have the results of a test with identical inputs. Overall, NovelAI turned to be slightly slower and requires more fuckery with prompt to make a better image. Sure it can also make more unique images with like unusual poses and bigger girls, but if you wish for consistency and able to be satisfied with just kinda curvy gals, I think Anything v3 will suit better
Oh and Anything is free to use. Here you can try it out with colab. https://colab.research.google.com/github/camenduru/stable-diffusion-webui-colab/blob/main/anything_3_webui_colab.ipynb
>>128661 Been meaning to try and give this a shot, but I spent more time trying to learn how to check for pickling than actually getting to download it in the first place. Never figured out how to set up the pickle checkers so I just moved on and went ahead with learning hypernetworking
>>128661 Is there a way to port anything over to a paperspace machine?
(7.36 MB 2816x4291 tmp2q9hsnwm.png)
Something that had never crossed my mind while fiddling with hypernetwork stuff is the strength adjustment, definitely was worth testing with the one I've been training
For stable diffusion web users, how do you guys keep navels from disappearing when you do Highres. Fix? It manages to create a navel in the first generation at 512x512, but when I either do img2img or highres. fix it'll always just remove the navel (2nd is img2img, 3rd is highres fix)
(1.98 MB 1280x832 doieayk4.png)
(1.36 MB 1024x1024 ojvefu3s.png)
My friend accidentally generated a sequence without realizing it. The AI created these two images on separate days when he wasn't trying to do that.
>>128898 its a good way to cater to whatever niche fetish you have without spending inordinate amounts of time sifting through shit on deviantart & twitter. and, honestly, its kinda fun >>128896 wish i had an answer. i also have a similar issue in which higher resolutions kill details like belly rolls (pic 2 & pic3). im not sure if there is a fix beyond just training our models more
>>128898 I don't do this everyday, I normally just set up the program to run while I'm doing something else that actually drains my stamina. I'll spend at most an hour just testing different prompts before I settle on something that seems nice, set it up to do something like 5 batches of 100, then run that before I leave home. I'd be miserable if I spent all of my time just fucking around with this, I'd rather be drawing by then, lmao
>>128898 At this point too I've also just become more impressed with seeing this thing at work. I still see it as more of a novelty than anything else, so I'm always a bit skeptical when people keep trying to say that this is going to be the death of actual artists. The only customers artists lose in the first place are the ones that don't have the actual money to buy commissions in the first place or the ones that act high and mighty thinking that DA points are still worth any actual value. Even with the artist I've been training a hypernetwork on it still can't get near the actual outputs that the artist makes, and that's perfectly fine by me. Only reason I trained it on certain artists was out of curiosity of seeing if that would make better results for fat or inflation stuff once fed back through img2img and taken off the hypernetwork or using less strength from it. That isn't to say I'm not annoyed by certain things from the "AI generated community." The massive influx of randos on Pixiv and Twitter creating fanbox or patreon accounts and claiming that they've put "so much time and effort into 'their work'" reminds me why I don't like AI art as much as others have. I'm just in it for the fun and novelty, and anyone trying to make money off of their $10 NovelAI subscription should be hit in the head with reality. After all, why would someone pay $10 to see "exclusive generated images" when they could just spend $10 on their own NovelAI subscription and make it themselves?
>>128661 Do I have to put this code somewhere? Or download everything by hand?
>>128356 >>128357 Muchly appreciated, my dude. Love the toned arms in some of them since they subtly imply strongfat. Hypothetically, if you or someone else were to make even more, whether even bigger, bursting out of clothes, or with whiter hair or darker skin, or all of the above, I certainly wouldn't be ungrateful, but this is a fine collection of work already.
What do you go through to get your images? Novelai?
>>128660 Damn wasn't expecting anything to be that good. That reminds me, I updated my colab to include more tools. https://colab.research.google.com/drive/1NVX7tldO72eKnMC4avJhdERZJblpcP-- I have the settings to match my drive setup, so scroll through and fix that for your needs. It will also only allow the downloading of some models within the colab itself. If you aren't making videos, skip the first setup section. And finally, read the instructions.
>>129500 No you just open this webpage, head to runtime tab and run all. It will give you a webUI link in a bit (based on gradio, you can't use IP because it's not your PC) and there you just use it as long as our overlords at google allow you. If the time is up, usually 3 hours of non-stop work, come back the next day or do the same thing as a different google account
>>129624 I mean you CAN run this code in your linux terminal by just removing all the ! and %, but you will obviously need a powerful linux PC with Nvidia GPU
>>127443 Where are you guys finding these models? I only have the old one much less the hyper ass or new furry one or whatever people are using.
>>129695 https://rentry. co/sdmodels It's also worth mentioning that as far as most people can tell, AnythingV3 is similar to the NovelAI leak, and it's not unlikely that it's just the NAI leak but with resumed training (and actually getting updated)
(335.49 KB 512x768 experiment (2).png)
(520.46 KB 512x768 experiment (5).png)
(423.29 KB 512x768 experiment (7).png)
(431.74 KB 512x768 experiment (1).png)
(499.57 KB 512x768 experiment (6).png)
>>128113 Trained a new model with a lot more steps and a new dataset. Workflow for people interested: *Used the instance prompt 'stout' as my concept *Used the class 'plump' with 2000 regularization images generated from the Anything V3 base model. *I used a 78 pictures of fatties from my favorite artists. The pics were nude and almost nude fatties for training. A lot of editing was involved to remove backgrounds and other things I don't want my model to learn. *Pics were mostly front view shots (pics with big bellies, breasts, and so on..). However, there were few back view shots (pics with big ass, back fat, leg fat, and so on..). *Used Shivam's Dreambooth Repo on Linux with a GTX 1080TI (12GB) *I trained the text encoder with a learning rate of 2e-6 *Used the 4GB pruned Anything V3 model during training *Made three models. Each were 2.5k, 5k, and 7.5k steps during training. 5k and 7.5k models had bad results due to overtraining. 2.5k steps was the good one (pics are from this model). Amount of training steps may vary for people, but I might try a learning rate of 1e-6 next time. Some tips for people interested in training their own models: * Know what you want (do you want a model that generates a fat ass or belly?) * Mixing ass and belly pics for my training images gave me weird outputs. * Be very careful with your training images (consider the objects in your training images such as clothes, poses, and special effects). You may want to simplify, edit, and organize your training images because Dreambooth has a habit to learn obscure things. * You may notice some sparkling effects in some pics from this post; this was a result from having training pics with sweat effects.
(446.57 KB 512x768 1.png)
(440.96 KB 512x768 3.png)
(279.32 KB 512x512 5.png)
(456.07 KB 512x768 2.png)
(439.37 KB 512x768 4.png)
>>130020 wow a GTX 1080TI. How much did that cost you?
Immobile catgirl with black hair and green eyes
Immobile catgirl with black hair and green eyes
>>130642 Jesse. Do your mom and dad know you're using the internet like this?
For Novelai, what did you set as an "Undesired Content" ? Tell me what settings are you using ?
>>130673 > Behavior I dislike will be rewarded anyway Although I am just as bad as you. I couldn't do immobile but here are wide fatties stuck in a doorway and busting out of a bodysuit
(203.98 KB 603x594 scale_1200.jpg)
>>130020 I've still yet to try and find a version of anything-v3 that is confirmed not to be maliciously pickled and I've spent a lot of time trying to set up a pickle checker but constantly running into issues compiling it or making it work. Do you have a link to it by chance?
>>130900 Is this not safe? https://huggingface.co/Linaqruf/anything-v3.0
>>122769 >>130020 The workflow is certainly helpful, especially the underlying model. I confess, I've not gotten anything nearly as good quality wise, even when I'm generating without training. With training, I've not been using regularization images, so perhaps that's part of it. I am curious if this could accurately replicate styles, or at least get closely approximate, since that's mostly where my interest lies at present. I've rather given up on trying to get the right sizes, just because it seems to give me massive distortions. I assume most people are using Anything V3? That seems to be somewhat better than WaifuDiffusion 1.3? What's most curious to me is that you're using 78 images, but only doing around 2.5-7.5k steps, which isn't too many epochs all things considered. Are you using gradient accumulation at all? I guess I'm curious this could be done using hypernetworks or Textual Inversion instead, since training a new model is pretty time and vram intensive.
>>130956 Yep, I was using gradient accumulation. Although, I'm not sure why 2.5k steps worked the best with 78 images. Should be more usually. I used to use NovelAI for fatties, but the AnythingV3 model gave me better results. I'm curious how well Waifu Diffusion 1.4 does with fatties though. I haven't touched hypernetworks or textual inversion yet. I should mention some other things: * I saw someone asking about where you get the Anything model and checking if it's safe. I got mine from sdgoldmine when it was released (4chan has a lot of links to that). Link that I used is probably dead. Automatic 1111 has a pickle checker, I think I used that (been a while). There's always a risk using these models, but I'm fine so far. * After I made the first model (2.5k steps), I changed the model path to that model in the training script for more training for another 2.5k steps and then again for the 5k and 7.5k step models. I didn't change the training or class images at all. This is what I did for continued training on Shivam's Dreambooth for Linux with WSL. I did not do any merging after I trained a model. * I'm currently reorganizing my training images for new specialized models. Before, my dataset had a mix of poses and sizes, but I am trying to seperate them into their own dedicated datasets to see if that works better. New more indepth workflow for people interested (general overview of what I'm doing): 1) Get your collection of fat pics. 2) Sort them. I made folders for different sizes (chubby, bbw, ssbbw, ussbbw, and immobile). Within each folder, I have different folders for different poses (standing, squatting, laying down or up, and so on...). I also like try to seperate poses into their own folder if they're strong front, side, or back view shots. Consider perspective as well. 3) After sorting, edit them if possible with photoshop or something. I reccomend removing the backgrounds all the time. You can choose to edit things in or out (adding a missing limb ot removing an accessory like a hat). Clone, blending, eraser, and blur tools in photo editing software are your best friend. 3.5) Crop your images to 512 by 512 (default) for training, but keep your OG images safe for future training you may want to do. You may want a seperate folder for your actual training dataset. You might do multiple crops for a single image. 4) Then, I train using the sorted images. For example, I use the folder of BBW, (front view) standing pose training images. I would use class images like 'plump' again or maybe 'plump 1girl standing' generated from the base AnythingV3 model. I would then use a instance prompt like 'bbw plump 1girl standing' or 'bbw_standing_frontview'. Learning rate should be 1e-6 (0.000001). You can use a higher learning rate like 2e-6 (what I used before, but overfitting may happen if you're doing alot of steps). Trained the text encoder as well. 4.5) Test the model with the usual settings (clip layer skip 2, cfg scale 12, euler, 20 steps). Use a prompt like this: masterpiece, solo, (your instance prompt) Also definitely use negative prompts to help your output. Use the usual negative prompts as you would with NovelAI. Wildcards are very handy for creating diverse fatties. Stay away from a pose wildcard or something that contradicts your concept. I got good outputs with these sizes. Width: 512 Height: 512 or 768 (both are good) 5) Some extra things: I train one concept at a time becaues I found it works the best + my poor GPU can't handle it. Training a single concept like a pose and size has its cons though. I am using the Linux version of Shavam's Dreambooth with 0.9.0 diffussers becaues it works for me. I recently tried using Shivam's Windows version of Dreambooth, but I got janky results (I heard there's a bug with the latest diffussers or something that causes degraded results). * If you are using the Linux version of Shivam's Dreambooth. You will probably have to convert the ckpt to diffusers. I used the latest conversion script in his repo. I converted the pruned one (4GB). *Formulas from the web. Steps = X of Image * 100 Class images = X of Image * 10 * Consider three things with your training images: Will Dreambooth be confuse by the clothing? I try to choose pics with not much clothes and more skin. Will Dreambooth be confuse by the pose? I might have pics of a girl holding her belly or bending down. I might want to seperate these complex poses into their own dataset. Will Dreambooth be confuse by the anatomy? If you're training with pics of multiple artists. Consider how they do anatomy with fat, you may find that an artist may do a stronger emphasis on details on ass or bellies. You may want to use training images that are similar and not too drastically different. Pick your training images wisely; this is key for a single concept Dreambooth model. You can spice up your model easily with a single image. I only reccomend mixing sizes and poses in your training dataset if it makes sense. For example, dedicated pics of a fat girl standing up with her arms up or down. For a scenario where it doesn't make sense, a full ass shot and a girl standing may confuse Dreambooth. I had one output where it made ass legs. This is one of the issues with training a single concept (why you should use multiple concepts during training). * If you can't find enough data for a certain concept, you may consider real pics of fatties or you could use 3D models. I am considering ass pics of real people for a dedicated fat ass model. * Bonus for people who have 3D fat girl models: You can create some interesting stuff by utilizing inpainting. Take pics of your fatty, and then refine the fat anatomy with inpainting. I had nice results using inpainting on fat girl pics from Koikatsu Sunshine. You can fix imperfections like jaggy bellies or clipping.
what is the name of the artist of this pic i cant make it out in the corner
>>131297 Look at what thread you're in
>>131297 I sometimes wonder how some of the posters here even function in the outside world...
(374.83 KB 512x512 1668834582047.png)
>>131297 Hi, is it cool if I mint this post as an NFT?
>>131320 You somehow managed to out-retard the retard, congrats.
>>131323 I'll take that as a yes. Now smile and say "NF cheese". 😎
>>122769 >>131050 This is all extremely helpful! Thank you! I do have one other question though. What sorts of images do you use for your regularization images? I've not really found much data on what those should be. For example, if I'm training a style, should I only use images in that style, or should I be using images of any kind that resemble the concept I'm looking for?
>>131551 I think it should be something close that resembles the concept. I generate images with similarities to the concept from the base model that I'm training. What other people do: I heard some people use images from the web as regularization images. They choose images that resembles their concept. I know someone who didn't use regularization images, but they got great results with alot of training. What I do for regularization images: For a model of big anime girls standing, I need regularization images of big anime girls or anime girls standing to help prevent overfitting. I generate the regularization images from the base model that I'm going to train on (AnthingV3) using a normal prompt like 'plump' or '1girl standing' or maybe something more specific like 'plump 1girl standing'. I didn't use any negative prompts when generating regularization images (I haven't messed around with that yet). If you're generating regularization images, I would choose a class prompt that the base model is familar with. For example, the plump class prompt is well known that can generate decent fatties. Although, the obese class prompt is not as well known; it can generate strange fatties. If some of your generated regularization images looks whacky and strange or black outputs, that's normal because that's what the model knows. I keep them for training. I heard that the weirder regularization images that your base model outputs is better for training apparently.
wow you guys, this is the best art there has ever been. The west has done it again! Moove over China and Japan. I'm going to go now to my special corner and play with my thingy (again)
(269.13 KB 448x448 1667243666996.png)
(261.75 KB 448x448 1667116013275.png)
(453.27 KB 512x512 1666501425681.png)
>>131592 Me too, thanks.
(99.73 KB 720x1325 Screenshot_20221221_1345352.jpg)
>>131324 Nope, you can't do that anymore. Look at this.
(3.06 MB 1920x1080 2022-05-26_2.png)
>>131642 A settlement is not an admission of guilt.
>>131642 Artists must response for their anti-creativity way against people, companies and progress
>>131644 Truth - it's literally a way to drop the suit, avoid the courts, and pay off the plaintiff with money sooner rather than later.
>>122769 >>131588 I see, so you generate your own regularization images. I hadn't considered that, because honestly I wasn't sure if they would help or hinder the training in general, since training a model is about trying to get something different out of it. I suppose i'll need to give that a shot. Honestly, dreambooth is so much different from training hypernetworks and textual inversion, it's kind of a struggle.
Everything you do is way too complex. I use the "Novelai" site, and now I can get good results. For example, if you select: "Nai Diffusion Furry (Beta V1)" Picture Resolution: Do not use "Portrait", because there is not enough space for the AI. "Square" for one person (1girl) "Lanscape" for two people (2girls) If you use "Landscape" for one person, there will be too many bugs. And for keywords I always use: {{{{{{masterpiece}}}}}}, best quality, {sweat}, {{{{1girl}}}}, {{plump}}, {{{{{thick thighs}}}}}, full body, necklace, earring, {{cleavage}}, {{{huge breasts}}}, {{{big puffy nipple}}}, {{{sweating profusely}}}, {{{wet clothes}}}, {{{{smelling}}}}, {{{{human}}}}, excited, elf's ear, freckle, {long hair}, {{{{{rolls of fat}}}}}, {{a worl of magic}}, {cellulite}, {{{{big ass}}}} To make the character a excited I use. : {sweat}, {{{sweating profusely}}}, {{{wet clothes}}}, {{{{smelling}}}}, excited It's to give an excited side. These are details, but so that the character is not neutral, and to give it personality I use: necklace, earring, elf's ear, freckle, {{long hair}, {{a worl of magic}}. And for body shape I use: {{plump}}, {{{{{thick thighs}}}}}, {{cleavage}}, {{{huge breasts}}}, {{{big puffy nipple}}}, {{{{{rolls of fat}}}}}, {cellulite}, {{{{big ass}}}}. The mandatory keywords for good quality are: {{{{{{masterpiece}}}}}}, best quality, full body. For the Nai Diffusion Furry the keyword "Human" is mandatory. And the {}, the more you put, the more the word is highlighted. And I very often have very good results. When I see that an image has potential, I use "Enhance". Either "Magnitude 5" when there are many defects or "Magnitude 3" for some retouching. It's up to you to test it, and tell me what you think! I know there are some who already knew this, I just wanted to share what I know. PS: Sorry I use a translator, so if you don't understand everything, it's normal ^^"
>>122769 How are weight gain sequences like these done? https://www.pixiv.net/en/artworks/103865982
>>122769 Have any furry model epochs besides the first one leaked yet? Which leaked model is better for giant boobs?
Yeah, how actually do you guys do sequences ?
How safe is this to do ? https://rentry.org/voldy GPU: NVIDIA GeForce RTX 3060 Laptop GPU CPU: Processor 11th Gen Intel(R) Core(TM) i7-11800H @ 2.30GHz, 2304 Mhz, 8 Core(s), 16 Logical Processor(s) Ram: 16 gb ram Hard Drive: 1 Tb SSD Are my specs. I'm just worried that if I do all of this it could damage my laptop somehow
>>132349 Should be absolutely fine, just watch your gpu temperature and don't do like 20 pics at a time or you'll overload your card either by memory or temperature
(1.31 MB 768x768 Gainshake_test_04.webm)
Might try to improve upon this style.
Hey, I don't have a GPU, and I'm too much of a schizo to keep logging into those sites... Can some kind soul ask for the AI to make some pic with belly button play? I know it's weird, but there isn't that much good art of it around. Thanks in advance
>>122769 Been looking around for different models, the PFG model seems to be able to do some really good bellies, though it requires a bit of cleanup to make good images at the end.
(429.03 KB 448x704 00449-3574385337.png)
(486.80 KB 448x704 00464-125611072.png)
(362.68 KB 448x704 00359-969831519.png)
(492.12 KB 448x704 00491-376245275.png)
(412.51 KB 448x704 00320-987234098.png)
(469.49 KB 448x704 00378-3114728569.png)
some Mikus. bless the guy who made the bad-image-v2 negative embedding
>>132747 PFG Model?
can u make more chubby/fat haruhis?
(472.82 KB 512x640 00019-1898991353.png)
(446.02 KB 512x640 00021-2018079115.png)
>>122769 >>132789 Yeah: https://civitai.com/models/1227/pfg Though most people seem to want to just use it to generate generic nudes, I've found it's got a lot of range in terms of poses and different kinds of softness. Nothing too huge without working with photoshop, and imo cleaning things up with photoshop and img2img produces better stuff, but it's perhaps one of the most flexible models I've worked with so far.
Does anyone have the blocked images from the deviantart of by visierah?
Looks like someone just ended up making a fatty model that y'all can mess with https://civitai.com/models/2664/bigger-girls-model
(598.85 KB 608x680 downloadfile.png)
(605.73 KB 608x680 downloadfile-2.png)
(596.48 KB 608x680 downloadfile-1.png)
(574.37 KB 608x680 downloadfile-3.png)
Here's the difference between it working on a prompt between SD in the first pic and then the new model for the rest. The prompt itself apparently doesn't even mention fat and it still produces these results in the fat girl model
It can effortlessly go big as well apparently
>>132880 Have you merged it with anything? I'm finding it tricky to get anime.
(585.00 KB 512x768 00695-627023404.png)
(493.57 KB 512x768 02182-1321580851.png)
(566.13 KB 512x768 02172-357742469.png)
(461.54 KB 512x768 00015-1578123926.png)
(479.29 KB 512x768 00023-2342219722.png)
(268.61 KB 512x512 00039-3534900500.png)
>>132877 it seems very good, but it also very often slips into a horrifying uncanny style. I also occasionally get speech bubbles as well as literal patron/twitter posts. But if you can dodge those, the results are great! here are a few good ones and a couple bad ones as examples (with bad-image-v2)
>>132877 While this is somewhat good on it's own, it really needs to be merged with another model. I merged it with Anythingv3 (anything 0.7 and biggergirlsmodel 0.3) Now that it's merged? My results are pretty good
>>132877 It’s sad I would need to make a whole separate account to make one for it for one download to test unless I’m misunderstanding this.
>>133037 could you post your merged model somewhere, or give a tutorial on how to merge models
https://sta.sh/2kjg5l1v695 Experiments with the bigger girls model, SD 1.4, and a merged one. Subjectively I like both the bigger girls model alone and the merged one even for non-anime style girls, but let me know what you think or if you want me to test this with other prompts
>>133037 To spare the trouble, would it be fine to share the model?
>>133090 how ya pulled this off
(412.11 KB 640x640 104066974_p0.png)
(434.73 KB 768x512 104105268_p6.png)
(508.16 KB 768x512 104105268_p0.png)
Alright, first time I tried writing this it just deleted itself, so I'll keep it short. There's a new thingy for anything.v3 that produces MASSIVE, more realistic ssbbw/immobile-sized results. "Borrowed" some images as examples from a Pixiv user Onyx Blade. They do mostly loli content, so spoiler tag enabled. Link to the model thing: https://civitai.com/models/2664/bigger-girls-model
>>133072 Seconding.
(4.07 MB 2944x1666 tmpiitzyh6x.png)
(797.51 KB 896x1576 tmplcn1im_l.png)
I'm back from my vacation and am finally catching up with developments/new things to try. The first things I did finally were try the anything-v3 from huggingface and that new bigger girls model (v1 so far since I saw some mentions of it being better for larger sizes than the newer versions). Before I got too deep into anything-v3 testing though, I went ahead and made a x/y plot featuring based pink cat girl vtuber just to see the differences between some of the checkpoints I have used before and some new ones using the bigger girls model (anything using that was merged with weighted sum A + 30% Bigger Girls V1. I also made a separate plot that's just anything-v3 since I forgot to add it into the x values the first time and did not want to just rerun it all for three more images on top of the fifteen there was already).
Unfortunately I was not able to get to merge the 1.2 beta from https://civitai.com/models/2664/bigger-girls-model because sd-webui stopped it saying that it was "malicious." I'm curious why the beta is malicious and not v1, but because I never could get the pickle inspector set up I don't know how to check if it's truly safe or not.
>>134155 Four, five, and six >>134157 Two and three Holy FUCK, Anon!!!
(291.75 KB 512x512 00004-1175332572.png)
(283.25 KB 512x512 00005-1604146252.png)
(290.94 KB 512x512 00006-3693380042.png)
(310.42 KB 512x512 00008-3125473651.png)
>>134155 Bigger Girls 1.2 merges really well for getting nice results with good variations of bellies. StrawberryMix + Bigger Girls 1.2 (.35)
who jerks off to this?
>>134155 >>134311 How'd you two get around sd-webui not using the 1.2 model? It still tells me that it's malicious when I tried again after redownloading it 24 hours from it's original upload
>>134313 I do. The other posters do. You do, too. That's why we post these. To gather more fap-fodder in a central place. Why are you asking pointless questions? Stop it, get some help.
i am having some problems getting the belly bigger than the tits even though i want both to be big. any suggestions to what i should write?
>>134351 I haven't used 1.2 yet but for 1.1 "average breasts" and multiple words for big belly (e.g. flabby, blobby, hyper belly, etc.) still tended to generate girls with bigger bellies than tits. YMMV
>>134316 I use colab for doing this whole thing because gods decided that amd doesn't need cuda cores in their life. Just took anything notepad from https://github.com/camenduru/stable-diffusion-webui-colab and added a !wget command, simillar to ones present: u just type the link for downloading bigger girls and type in location for in after -O. Everything else is pretty much same process, go to checkpoint merger, choose to what model u want to add and how much, type in the name for it and wait for like 20 seconds, then choose merged cptk on top left
(350.62 KB 4000x1800 photo128471823.jpg)
I think I've found one of great usecases even for not upscaled images. Just use em as your forgotten handheld wallpaper. It's literally perfect resolution for the fast generating kind and somehow looks so much better than on PC screen bruh
>>134407 I'm struggling to find where those models are. I think I understand how/where to implement the wget on my instance on HuggingFace, but I can't find StrawberryMix, BiggerGirls, or anything not listed here: https://github.com/camenduru/stable-diffusion-webui-colab Could you point me in the right direction?
>>134470 StrawberryMix is a merged model using already existing model checkpoints, so it's something you have to make yourself. BiggerGirls is something that's been linked a couple of times already and just requires some reading through posts to find it.
Man, this shit is fucking bonkers... Not even half trying getting this kind of stuff... is insane.
Some of my favorites with the model mixing described by >>134311. No img2img or photoshop, but I did upscale the middle one. On a more critical note, have people noticed that faces start getting samey if you use the model overlays with the bigger girls model than if you use the models alone? Especially if you use the prompt to force them to be a certain style and hair/skin color? Even relatively common hair styles (long brown, shoulder-length blonde, blonde ponytail, etc.) seem to consistently get similar faces with the prompts I'm using