/bbwai/

(1.7 MB, 1024x1024, 00050-675933285.png)
I thought this could be a good place for users to post general questions and hopefully get some answers from big brain anons.

I'll start.

When generating images using wildcards in web-ui, is there a way to determine which option from the wildcard term was selected to generate an image?

I feel like being able to know this would be useful for fine-tuning a prompt.
(795 KB, 896x896, img2img.png) (498 KB, 1675x957, promptsettings.png) (178 KB, 1454x361, pnginfotab.png)
>>2561 (OP)
Great thread idea and an easy first Q to boot.
By default, all image gens from the WebUI are embedded with their final prompt. Simply select the image after generation and the full generation settings will be displayed in the area below the generations, or drag the image to the PNG tab if you want to look at the prompt info for an older generation that's no longer in the webui.

In the posted example, I used the body-heavy, size, and negative wildcards, but by selecting one of the images after generation, I can see that squat was used for body-heavy, vast was used for size, and jeering was used for negative.

A reminder that we strip exif data here for anonymity's sake, so you will not be able to use the PNG info tab on images posted here, unfortunately.
Is anybody else having problems with Colab? My models were working fine with it a few days ago, but today they suddenly were all throwing "modules.devices.NansException: A tensor with all NaNs was produced in Unet" errors.
>>2614
Recent SD update broke lots of things. For the time being you can try adding "--no-half --no-half-vae" to your launch arguments or use an earlier commit. Although I have no idea how to do either of those since I don't use the colab, maybe another anon can follow up
(1.8 MB, 1024x1024, 00011-2637848699.png)
>>2563
Thanks for the help!

Has anyone found decent ways to generate food in images that doesn't look all wonky most of the time?

I won't even begin with trying to portray the subject eating, but I've managed to get some good generations with the subject holding food...sometimes.

Positive prompt is:

((holding food))

Negative prompt is:
(((food not in hand, food out of hand, food in legs, food in belly, food in face, food on ground, food in floor, deformed food , bad food , food on thighs, oversize food, too big food, weird food, foodin thighs, food on butt, food on belly, food coming out of skin)))

food is a wildcard for your own list of foods.
(1.0 MB, 1024x1024, 37673-2975736378-(masterpiece, best quality, highres, detailed_1.1), mori calliope, red eyes, pink hair, tiara, hooters, orange shorts, white tan.png) (836 KB, 1024x1024, 11265-3204082646-masterpiece, best quality, highres, latex bodysuit, cat ears, in a dungeon, incredibly wide hips, huge belly, hands on belly, si.png) (1.0 MB, 1024x1024, 24066-818839546-masterpiece, best quality, highres, _lora_allVtubersLora_hll31_1_, ookami mio, orange eyes, hair ornament, white t-shirt, cleava.png) (989 KB, 1024x1280, 30485-4032812664-(masterpiece, best quality, highres, detailed_1.1), _lora_allVtubersLora_hll31_1_, nanashi mumei, hololive, brown hair, brown ey.png) (2.0 MB, 1280x1024, 32056-3337278202-(masterpiece, best quality, highres, detailed_1.1), _lora_allVtubersLora_hll31_1_, nanashi mumei, hololive, brown hair, brown ey.png) (1.1 MB, 1024x1024, 25046-3049192058-masterpiece, best quality, highres, _lora_allVtubersLora_hll31_1_, ookami mio, orange eyes, hair ornament, crop top, sweatpants,.png)
>>2687
You might be able to clean up your pics with inpaint (>>7 (Cross-thread)) but besides that there's really not much you can do. Picrel are the only times I've managed to make food look decent and they still took a shit ton of inpainting

I would imagine it's an artifact of how these models are trained- they're fed 512x512 images, and while that's fine enough for shaping a girl, it isn't enough to capture the detail of how she holds food. And there's little to no training on food itself (eg just a picture of a burger, or a burger being held by a hand), since the girls are supposed to be the focus. Maybe at some point someone will release a lora or something that can do food better.
having trouble achieving body contrast between characters in the same image, like how would you structure a text prompt for a fat female in doggy position with a muscular male, for example?
>>2742
I would use latent couple, possibly with control net for the posing (using open pose skeleton). You can then describe each part of the image as normal.

Without latent couple, it'll just get mixed up with the tags.
Did this site lose all the metadata from the PNGs or have people just been stripping it manually? I swear I used to be able to pull some images and extract tags it really helped in trying to get a baseline to seed new ones.
In a prompt, how do items
without brackets,
(with brackets),
(with brackets and a weight:1.2)
compare? Are the first two comparable to a specific weight?
>>2808
no parenthesis are normal weights
((parenthesis)) are the same as (parenthesis with weights:1.2), with every parenthesis being roughly equal to 0.1 weight. the numerical weights are better because they're easier to read, edit, and make (highlight something then do ctrl+shift+up/down to control weight)

[brackets] and {braces} don't do anything. they still do on NovelAI's website, but on stable diffusion webui, they're useless. <arrows> only activate loras, like <lora:example:1.1>

>>2791
bbwchan doesn't do metadata anymore, yeah. Unfortunate, but probably for the best. Just ask for catboxes, they contain metadata and you can put them in png info like normal
I was unable to find any web options like collab notebooks (.ipynb) mentioned in any of the threads, unless I am blind. What is the way some of you go around using it?
The basic guide mentions using artist names to influence style, but is this dependent on whether the checkpoint or lora you're using recognizes the artist?

Are booru tags special in some way? Would they be distinct from writing the same words without underscores?
(409 KB, 512x512, 1.png) (312 KB, 512x512, 2.png) (272 KB, 512x512, 3.png) (330 KB, 512x512, 4.png)
Would like some help for combining images with latent couple

I generated images 1 and 2 using the same model, each with their own prompts. Then, I used GIMP to superimpose them in the positions I was aiming for (male behind female), as seen in image 3. Finally, I used the rectangular region mode for latent couple, defined 2 rectangles for each character's region, both with weights 0.8, used the same generation prompts separated by AND, but kept getting images that missed the mark/flipped genders/look overcooked/etc, like image 4. My parameters were CFG 10, denoising strength 0.7, DPM++ 2S a Karras at 20 steps. Am I doing something wrong?
(231 KB, 2048x1281, tags.jpg)
>>2885
> The basic guide mentions using artist names to influence style, but is this dependent on whether the checkpoint or lora you're using recognizes the artist?
Yes/no

Stable diffusion and by extent all of its checkpoints have training on a few notable artists. https://stablediffusion.fr/artists
This style training is pretty rudimentary and also gets muddied down further by the anime/fat/fetish/whatever training that the checkpoint goes through. But, yes, all checkpoints by extent recognize these artists at least a little bit and change their outputs accordingly (even if the output does not resemble the artists' art at all)
You can leverage these pretrained artists to get more varied fats and pics (>>2538 (Cross-thread)) but I don't really know why you would do that over using style loras or specific models/merges or better prompts.

And of course only mainstream artists are recognized. Someone like Kip or Trinity will never be in SD and you'll have to use a lora or model specifically trained on them

>>2885
> Are booru tags special in some way?
Yes.
The vast majority of models are trained using tags for simple, standardized descriptions. Particularly, booru tags (danbooru, gelbooru, rule34, realbooru, etc) are used because they're pretty accurate and it saves model makers a lot of time collecting & tagging things themselves. So, those models will understand booru tags much better than if you were to try to make something up. For example, a model is likely to understand the tags "paizuri, breast lift, NSFW, penis, pov" more than "titfuck with girl squeezing her breasts". The tags to use depend on what lora and model you use, since some models may be trained differently than others.

> Would they be distinct from writing the same words without underscores?
No
Underscores are treated the same as spaces for all purposes including prompting and training

This extension helps keep track of tags because it is pretty annoying to remember them or look them up. Pretty useful:
https://github.com/DominikDoom/a1111-sd-webui-tagcomplete
(443 KB, 768x512, 38237-641948082-masterpiece best quality highres detailed_1.1 1boy 1girl handjob NSFW penis_AND masterpiece best quality highres d.png) (24 KB, 1021x82, Screenshot_20230405_220053.png) (1.4 MB, 1408x1024, 38806-1441537196-masterpiece best quality highres detailed_1.1 in a restaurant 2girls fat gigantic belly wide hips_1.2_AND masterpie.png) (32 KB, 1400x88, Screenshot_20230405_220434.png)
>>2891
Post a catbox?

> each with their own prompts
Your prompt is probably no good.
Every line is supposed to have the same prompt and the only thing different should be the character.
So if I wanted a handjob I would use "1boy, 1girl, handjob, NSFW, penis" in every line and the only thing that would be different is the character prompt. Or you wouldn't even use a character prompt and SD will just fill it with some generic moeblob
You could also try using controlnet. I would use canny or depth since you already have an image that's roughly what you want but you could also use openpose
https://github.com/lllyasviel/ControlNet
https://github.com/fkunn1326/openpose-editor

>>2815
Bbwchan doesn't have any colabs, you'd have to use the /vtai/ colabs (>>9 (Cross-thread))
> Cloud Hosted Install
> Colab: https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast_stable_diffusion_AUTOMATIC1111.ipynb | Paperspace: https://rentry.org/865dy
> Colab for Complete Retards: https://colab.research.google.com/drive/1STL60qfoY-iSlhRb9zFETRLTqhNbznRf

I think the retards' colab has a few fat models preinstalled? I haven't used it, and even if it didn't, just upload one yourself
(405 KB, 512x512, 1.png) (283 KB, 512x512, 2.png) (228 KB, 512x512, 3.png) (278 KB, 512x512, 4.png) (267 KB, 1853x774, LC overall prompt.png) (124 KB, 1773x1070, LC region settings.png)
>>2894
Thanks for the reply. The previous prompts were inconsistent, so following your advice, I redid them to ensure that they lined up now, but the latent couple results still look overcooked & have some traits mixed up?

1 (man): https://files.catbox.moe/d8yrhp.png
2 (woman): https://files.catbox.moe/nzejwi.png
3 (man + woman): https://files.catbox.moe/syp4x8.png
4 (latent couple result): https://files.catbox.moe/avxew5.png

The model I'm using is a 80/20 merge of Bigger Girls V2 and Uber Real Porn Merge, if that's relevant
(357 KB, 512x512, 46963-3121099229-(masterpiece, best quality, highres, detailed_1.1), 1boy, 1girl, boy with red hair, girl with blue hair, sex, penis, NSFW, doggy.png) (387 KB, 512x512, 46954-3380935075-(masterpiece, best quality, highres, detailed_1.1), 1boy, 1girl, boy with red hair, girl with blue hair, sex, penis, NSFW, doggy.png) (374 KB, 512x512, 46953-502193223-(masterpiece, best quality, highres, detailed_1.1), 1boy, 1girl, boy with red hair, girl with blue hair, sex, penis, NSFW, doggy.png) (41 KB, 512x512, tmpfj2p9clu.png) (356 KB, 512x512, 46951-1293582204-(masterpiece, best quality, highres, detailed_1.1), 1boy, 1girl, boy with red hair, girl with blue hair, sex, penis, NSFW, doggy.png) (356 KB, 512x512, 46950-3549808952-(masterpiece, best quality, highres, detailed_1.1), 1boy, 1girl, boy with red hair, girl with blue hair, sex, penis, NSFW, doggy.png)
>>2908
Definitely looks a lot better than the last pic, good job
> traits mixed up
Latent couple is mixing things up, you can try matching your step count with the 'end at this step' count.

> overcooked
May have to do with the size of your prompts, I get the same thing when I use lots of tags. You can try removing some less important details from your negative&positive prompts, and/or turning down your cfg scale- I use 4 for complicated prompts

You could also try getting rid of some of the pose prompts- 'on all fours' might not be necessary since you already have 'doggystyle, bedroom', and you don't need 'full body' since it's not a full body if he's behind a girl.

But I did find that making the poses ahead of time with openpose made it much easier to work with, so seriously consider giving that a try (all pics same prompt, pics 1-3 used the pose, 5&6 didn't)

Besides that I'm out of ideas, I'm not great at using latent couple. Maybe another anon can help you further
>>2931

alright, will try trimming prompts and openpose when i can, thanks for the help!
>>2687
The best solution is to try and find a model that does food well and merge it in.
I have had good results with some of my merges without needing to inpaint at all.
Does anyone know what prompt to use to make a Yiga Clan outfit from BOTW in Stable Diffusion
>>2931 Is that a girl with balls getting fucked inbthe anus by a wooden table?
Thank you all for your patience with me in the other thread. My last post was about finding more AI art? Is there a site for AI art like a booru or something with only AI generated art?
>>3337
Does anyone know the correct prompt for this outfit?
What is the best model for non-anime art? Do the normal bbw models work for different styles, like video game or 3d art?
Has anyone gotten any local chat AIs to work?
>>3453
I don't think anyone has made a photorealism merge for bbw-step-30000k or Biggergirlsv3. Most of the fatty models are trained on artwork, and then merged with anime models. Maybe check the realism thread?
(750 KB, 1800x1079, 031.png)
When trying to train a lora for novelai base, are you supposed to escape brackets when tagging?
Do you tag it as 'rhea (fire emblem)' or 'rhea /(fire emblem/)' in the caption file?
>>3638
You tag it with the escapes, and with underscores, so you'd use "rhea_/(fire emblem/)"
Is it dangerous to use chatgpt for fat fetish wg fiction roleplaying? Because I have gotten really great results, it's all I've been using. Just wondering if it's safe or not, because of course every other prompt has inappropriate guideline warnings.
>>3747
the worse they'll do is ban you, and even then they probably wont
>>3748 thanks buddy :)
>>3747
I've been repeatedly demanding that the chatgpt domain and mobile app give me slight variations of some chick eating way too much at a buffet.

For some reason if i do this with openAI they'll call me fatphobic, but the chatgpt domain and the mobile app don't even care. I'll even ask them to describe the genre of the story and they'll say "comedy / obesity fetishization"

Skynet is onto me...
Does anyone actually use "Fix Faces"? Feels too general to reliably use and would make things more complicated and rough than would be worth it.
(467 KB, 3456x1382, rf.jpg) (949 KB, 1592x598, 63cc2f7ff604a3d2137f7e02_Untitled.png)
>>5075
Face restore is only for realistic models, on the stylized/anime models it results in uncanny hair and eyes and mouths.
On realistic models it still doesn't work very well but you're also supposed to configure it ahead of time.
Either way you should probably just inpaint >>7 (Cross-thread)
My day be so fine, then
>ValueError: images do not match

lads it's divisible by 8, why is this so off-and-on
>>5211
Yeah, the inpaint guide says to only use 'only masked' for faces, but you're good to use it and the relevant settings pretty much all the time. It should do away with that error, and usually gives better results, while being faster than 'whole image'
>>5212
One would think, but this happens most often with 'only masked'. Sometimes I have to launder it through img2img to do away with the error, even though I inpainted it by copying from txt2img>inpaint straight away.
are there any free sites for ai that allow fat prompts? even if it's not that good or takes a while? computer is amd and it just takes ages making any images, so i cant even play around with prompts
>>5246
The google colabs were a thing but they're starting to shut down free usage.
The paid plans are still really good though. If you pay $10 for pro you get ~50 hours of generation time a month. But many anons report having access to the gpu far, far, far after they've run out of credits. It's a priority system and paying gives you REALLY high priority even without credits.

Paperspace also has a free plan but not sure how that works.

There's also stable horde https://stablehorde.net/ but the wait times are horrible and you don't have access to fat models

> amd
What's your card? Your launch arguments? Amd gets shit on by nvidia but anything modern should still be usable
>>5249
Radeon RX 570 Series.
Not sure exactly what my launch arguments are,
--precision full --no-half
it's about 10-20 minutes per image depending on settings
>>5249
i would appreciate help on where to find the launch arguments and where to get better ones lmao. i have 32GB of physical RAM surely that can help
>is fucking clueless
>>5250
> Radeon RX 570 Series
my condolences
> 10-20 minutes per image
I got similar times (on an rx6800) and chalked it up to AMD support being shit. Then I looked at task manager and realized it wasn't using my gpu at all, and the sd console was bitching at me about not recognizing my gpu.
Maybe it's the same thing for you.
https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/1220
I don't remember the specifics because I switched to a 3080 a while back but I distinctly remember messing with the torch command and launch arguments eventually got me down to a much more reasonable 3min/img

>>5251
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Command-Line-Arguments-and-Settings
> The recommended way to customize how the program is run is editing webui-user.bat (Windows) and webui-user.sh (Linux)
I would add --medvram and --opt-split-attention to your launch args, reduces memory usage by a lot at the cost of ~15% speed so you can generate much larger resolutions.
> i have 32GB of physical RAM surely that can help
all that would do is let you cache models for faster switching. SD cares about gpu and gpu only
Attempting to follow along with the guide, but I get stuck at the inpainting stage. After I send the resized image to inpaint, no matter what I click, I always get the "ValueError: images do not match" error.

Also, I can't find the Upscaler option anywhere on the screen. Any help?
anyone got reposts of all those fat loras that were shared around? like the trinityfate stuff, the jaykuma stuff, and the mysterydad stuff?
super random, but anyone know which prompts to use to avoid a split belly? like a pregnant-looking one, or one where it's split horizontally
>>7250
inflation, stuffed, ready to pop, in positives
flabby, double belly, fat rolls, belly rolls, in negatives
is there a list of general free image gen services? like frosting.ai? even if there's little to no control, just to have many alternatives. if there's no list, feel free to share as many as you know of. thank you.
>>7630
Stufferdb has a new, free image generator that lets you choose from a bunch of models and loras. It's pretty shitty though, it only outputs up to 640*640
There's also webui for google colab, and you can find a few presets that come with models pre loaded, but I think that's been nerfed for free use?
Dall-e is neat but doesn't do any fats
Pixai.art is a thing and there's a thread for it but don't know much about it

The best solution is to just run it on your own pc, and if you can't, to use the google colab. The rates are much better than novelai and you can use whatever model you want
As someone who considers themselves 'technologically illiterate', but is also interested in all of this, where would I even start?

What would be the easiest and best program to use? Would it be better to try to create a more realistic character or one in a more artistic style? How easy or long of a process is this? Is it just typing in a bunch of prompts like a chat ai or is there actual coding to it?

Is there even a point in me trying to start this process or would I be too far lost in knowing how to even create an image?
how to resolve this error:
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument index in method wrapper__index_select)

happens whenever I select a VAE, apply settings, and try to generate.
Anyone have some tips/advice on how to start creating anime bbw AI art? Do I use Stable Diffusion and or what software/program do I use? Hopefully this isn’t a stupid question.
>>2561 (OP)
How do you keep your openpose images organized? Is there an easy way to recall which pose you used for a particular image?
>>7922 >>9506
The guide thread >>2 (Cross-thread) is a pretty good place to start and covers the basics in regards to installation and prompting and some FAQs

> creating bbw anime art
Loads of options.. Once you got stable diffusion webui setup, the simplest is to download a bbw model (I recommend aurora borealis https://huggingface.co/mariaWitch/ExperimentalBiggerMerges/tree/main) and just start asking the ai for fat girls.

https://civitai.com/ Browse civitai for loras for different concepts, like your favorite character or art style or fetish or whatever

Once you get more advanced you can start working with fat loras & different models like symmix or bgv5ex or biggergirlsv5 but learn the basics on aurora or GAME first

> creating real bbw art
I don't know much about that but I believe the best bet is biggergirls' realism models https://civitai.com/models/2664?modelVersionId=57548. If that doesn't float your boat you can also merge it into other realism models or use fat loras on normal models

>>9608
> Is there an easy way to recall which pose you used for a particular image?
Put your image into controlnet and select a preprocessor and it'll use the pose in the provided image in your generation, there isn't really a need to keep a pose library if you already save your images
If you use the preview function you can get the pose without generating anything if you want to save it
Potentially dumb question: What would be a good way to make a prompt produce a "western" art style as seen in some of deviantart and twitter's better offerings, rather than anime styled? For some reason I just prefer the simpler body shapes, shading and faces seen in a western comic/cartoon art style, but actually getting the AI to understand what that means and produce fitting outputs that aren't anime/manga is very hit-or-miss in my efforts.
>>9784
Easiest solution is to use a different model. There are two models I go to if I want a look that isn't anime.

The first is called "TFM American Cartoons Model" which doesn't really have a style one could point at and say "It looks like this," instead it's more about colours and lighting but is undeniably western in it's results.

The second I use is called ManMaru Mix. This one is named after its purpose of generating characters with 'round eyes'. It's far more anime-y than other models that are available, but I adore its colours and lighting. In the time I've used it, it always reminded me of comics I'd find on Twitter.

Another option would be to use a model like 'AnyLora' or 'Something' alongside a character LoRA from a cartoon of your choice. These kinds of models let the style of the LoRAs come through really well and can do most of the work for you.

In all, I guess you should just go onto Civitai, filter by Checkpoint and starting scrolling until you find something you like. You're trying to find something specific, so you'll have to look for it.

TFM American Cartoons Model: https://civitai.com/models/25627?modelVersionId=38904
ManMaru Mix: https://civitai.com/models/86277/manmaru-mix
>>2561 (OP)
1) How hard is it to train LoRas, specifically for a set of artists?
2) Are there any model or weights that are specifically trained on fat women? Or pregnant women, or inflated (circled belly) women?
has anyone tried using GirlfriendGPT as an alternative to character ai bots? I can't quite figure out how to set it up but I don't think it is nsfw limited
I've noticed that some of the bots on venus.chub have alternate greetings, and I'd like to try and incorporate those into some of my own. Anyone know how to?
>>10499
You can give characters created on chub.ai (not venus.chub.ai) alternative greetings if that's what you're asking
>>10136
I have not attempted it yet, but looking at LoRas using safetensors_util.py to view their metadata (which ALSO gives you insight into what tags work with the LoRa) I see some of the mistakes they are making. For example, take "Obese Girls | Concept" https://civitai.com/models/48197?modelVersionId=52816 (a.k.a. <lora:Fat:1>) This seems to have been trained on 1079 different images if I am understanding the metadata correctly. This is very achievable. (You might possibly even enlist Amazon Mechanical Turk, although that might violate their terms of service. I am going to look into it. You could maybe feign being part of an internet safety group.)
Anyway, back to that LoRa: I asked it to generate 50 images with: "(weightimmobile:1.2),<lora:Fat:1>". All 50 were belly shots. My guess is that the artwork for weightimmobile was biased the way it is on deviant art. Few artists are devoted to ass shots of their immobile beauties. (And I heartily thank each and every one of them.) Even though I next circumvented that bias, this is a noteworthy rule: Do not claim a whole category with a LoRa, and then fail to represent complete aspects of that category. So I got around this by cracking open the metadata for the LoRa using safetensors_util.py. In there were all the terms used. I found two that were ass related: "ass", and "huge ass". I searched for synonyms, but there were none. I then updated my prompt to be: "(huge ass),(ass),(weightimmobile:1.2),<lora:Fat:1>". I got back 50 ass shots. :-) There were about 5 great ones, but most had issues. Buttholes were exposed like belly buttons and positioned so, as if on their tailbones, tailbones had umbilical hernias, sometimes the whole backside was just a belly. The other things that I noticed in both runs, were things like weighing scales, and a gym deadlift station showing up. The only place you will see the term "scale" show up in the metadata is "ss_bucket_no_upscale":"False", lol. There is also nothing gym related except "gym uniform". Those portions of the image that contain detailed info have to be identified in the tags, because that is a lot of information to diffuse into the model. By providing the additional tags for high information pieces of your training images, you can prevent this.
The information in the tags has to match the information in the image, and be equivalent in terms of the amount of information it conveys, or it will have these kinds of unasked-for distortions.
picrel (of woman with gaping belly-button back side) omitted, lol
>>11013
That said, and thank you for coming to my TED talk, lora:Fat was trained on deviantart anime, and I was able to render photorealistic images from it, so I think that one could create a LoRa with the crudest of images, or the img2img re-renderings of those crude images, and develop a powerful tool for a well known style based on that.
Hey, there was a Character.ai chatbot of "Wanda" from "The Fairly OddParents", BUT I can't find it anymore, help me find it? PLEASE
>>11013
> I needed to specify that I wanted an ass shot to get an ass shot.
> I discovered this by prompting for 50 images without tags for an ass shot and not getting a single one, along with fumbling around in the data of the lora to find the words the lora was trained on, instead of just prompting for an ass shot
are you serious?

> prompting for 'gym uniform' would introduce other things unprompted, like gym equipment, gym background, or weighing scales
Yes, it's every model's way of making the scene more cohesive. Prompting for 'gym uniform' or other athletic wear will often make the setting a gym as well, the same way prompting for a bikini will sometimes put the girl at a beach or in a pool unprompted
You can get around this by specifying the setting you want and/or putting what you don't want in negatives.
So, if I wanted a girl in a bikini, but didn't want her at the beach, I would specify a setting by adding 'indoors, living room' to the prompt, or "outdoors, beach, water" to the negative prompt.
And for you, if you don't want gym equipment or scales, put that stuff in the negative prompt

> the lora doesn't have very much training on different poses and body types, makes it hard to get what I want
Valid criticism, assuming you know what you're doing. Which you probably don't, tbh

>>11015
> you can make a lora from crude images
You can make a lora from basically anything, and like you saw, they're pretty flexible
> thank you for coming to my TED talk
put it in the general next time, this thread is for questions, not your 'research'

>>11016
??
I'm quite new to generating fats via ai. Could anyone give some sort of starter tips? Or.. basics?
I am trying to create weight gain sequences with stable diffusion. How do I make it so the face stays recognizable as the same person between generations?

Right now I am using the inpaint tool to mask the face, but I want some variation of expression, face fatness, angle, etc. The inpaint tool also tends to lead to an unwanted cut and paste sort of look too.
>>11408
>>2 (Cross-thread) goes over the basics and setup
here's some stuff a lot of newbies miss, some of which isn't there:

> 'face restore' option is only for realistic human faces, and not for anything anime, furry, or otherwise stylized. it WILL ruin faces
> if your image is desaturated, has weird texturing, and/or purple splotches, you need to download and enable a VAE
> disable image previews- they slow your generations down, and aren't that helpful anyway unless you upscale EVERYTHING or have really long gen times
> best way to generate is to make a big batch of low res gens, and pick your favorite(s) to upscale, to save time.
>> use 0.5-0.55 denoise, resrgan anime upscaler, x2 upscale- these are good highresfix settings most of the time. latent is ok sometimes but tends come out 'fuzzy'
> normal models (such as those on most free gen services) can't do fat, you will need to download a model or lora that can do fat, like biggergirlsv5 or symix
> stufferdb has an ok free gen service with fat models preloaded
> if you frequently run into out-of-memory errors, look into enabling xformers if you're on nvidia, and puting --medvram in your launch args
>> look over all of the launch args to see if anything interests you, and google the ones that do that you don't understand https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Command-Line-Arguments-and-Settings
> please god learn to inpaint
> you will learn more useful skills by experimenting with the settings and prompts than by brainlessly following a guide. you've got nothing to lose except your time
> r*ddit's stable diffusion sub is actually a really good resource for finding news, new models, loras, techniques, twitter-tier gossip, etc
> the stable diffusion threads on 4chan are also good for that, especially /vtai/ on the 4chan vtuber board
> taking requests is a good way to improve quickly. you'll learn lots as you try to do stuff you otherwise haven't and wouldn't have tried before

It's also good to know that people love geeking out about their hobbies and most will gladly sperg out for anyone that'll listen, like I just did.
Ask nicely and people will tell you anything. Discord, 4chan, bbwchan, reddit, whatever, wherever,
Hope that helps

>>11511
> How do I make it so the face stays recognizable as the same person between generations?
Not really possible for generic characters/OCs, since there isn't a way to prompt for facial structure. All you can do is prompt ethnicity/race, obvious facial features like eye color, scars, makeup, and maybe celebrities, if your model understands them
It gets much easier if you use a lora for an existing character/person, though

> The inpaint tool also tends to lead to an unwanted cut and paste sort of look too
Use a low-mid denoise (0.45-0.55), 'inpaint area: masked only', and a resolution of 512x512 - 1024x1024, and a padding of 256
Haven't done any generations in about a month, what is the current go to for models? I think I was last using bbw_hll3.1 final.
These are probably stupid beginner questions since I just set up SD this evening, but what are the go tos to eliminate bad anatomy?
>>11976
Latent or r-esrgan upscaling at 1.75x+ does wonders for shit faces and hands. Also the easy negative text inversion seems good but I did fine without it too. Rip your gen time tho.
>>11661
Bgv5ex lora and a model that fits the style works better than aurora and other bbw specific models IMO. Perfectdeliberate has an anime and realistic versions that seem good with this lora. Symix lora is picky about prompts and model but bgv5 lora seems more consistently good. Also control net with a fat girl as input can wrangle all manner of models into making fats.
Been playing around with it a bit and had better luck but I can't get in painting to work which is an issue. Either does nothing, runtime error not enough memory or produces a smudge. GPU is AMD Radeon RX 6800S, 8GB VRAM. Are the crashes just my GPU not having enough memory to do this sufficiently or did I set up something wrong?
>>12046
Not experienced enough with inpainting to help with that, but fun fact I've noticed: when I trigger an out of memory error the gpu memory doesn't get freed until I restart the webui.sh. Might help to have a monitoring tool open to make sure it's freeing gpu memory correctly. Also you can always copy the models folder and reinstall 1111.
i love bbw
me too
Im new here
One question, does anyone who has paid for the spicychat service know if it is worth it?
Is there any way to generate ai art without a super computer or some paid subscription service?
Someone knows how to make body inflation videos at Gen-2?
I'm new to AI creation and a bit confused about trigger words. I've been trying to create/publish LoRas with SeaArt AI, but that section confuses me.

Also, how does the SeaArt AI tool compare to other AI tools?
>>2561 (OP)
I would like to request a website change: a thread bump limit of 300.
(151 KB, 1024x1024, Peach Burp AI.jpg) (211 KB, 1024x1024, 2B Burp AI.jpg)
I found some gens on 4chan with burps in them and I'd like to ask how one recreates the open mouth and text bubble burps. I assume the latter is due to editing the gen itself.
Anyone got a decent guide as to how to set up a lora based on a specific person? I was hoping to set up a Juicy Jackie lora so I can recreate here in AI and possibly make a version of here with big fake tits or make other adjustments like make her Asian etc.
>>13614
Personally, I used this guide to learn how to make LoRAs: https://civitai.com/models/22530?modelVersionId=26903.

Though in this case, you'll need to gather a large amount of images without the help of an image scraper. You *could* use as few as 10 to 20 images, but that will result in a low quality LoRA. The more images the better, but it'll increase the time it takes to create the LoRA.
>>13628
Thanks, I'm setting up a cloud computing solution so I can hopefully run it for as long as it takes. I'll collect a ton of images andm hopefully it will work.
Is There A Actual Discord Server For This, I Want To Improve My Results
someone knows how i can make an inflation video at gen-2?
I have a question, how would I get it to make a bunch of images where the character looks the same but either bigger or in a different position>
I was trying to make a buff Bowsette belly inflation video using Gen-1 but it doesnt generate Bowsette characteristics even if i put them on the prompt and doesnt show up the belly meanwhile the base does. So i would like to know a hint to do the video
(4.7 MB, 4624x2604, IMG_20231020_214915.jpg)
I want to generate a bbw in Bing ai,but he doesn't want to show me the result. What to do?
I'm gonna miss bbwai more than any other boards. I say that as someone who has been coming to bbw-chan in its various forms for 20 years.

I don't have to explain to anyone here how the fat fetish community at large doesn't really like AI-generated stuff.

Where will we go next? Any places we're gonna congregate to post fat AI images?
(101 KB, 1280x1280, a_little_heavy___by_ketbiwankenobi_dg9qiru-fullview.jpg) (117 KB, 894x894, do_you_still_think_i_have_a_big_ass___by_ketbiwankenobi_dgcm9ct-pre.jpg) (355 KB, 1920x1920, you_lose_by_ketbiwankenobi_dgags5i-fullview.jpg) (1.5 MB, 1200x1200, dg6459v-ca17ba8a-03a5-4d8a-907c-4214fe29c070.png)
which of these AI sites/programs would be best suited to try to generate facesitting pictures like the ones in my post (these are not mine, taken from DA)?
Someone knows Runway gen-2 prompts for make a good inflation video?
I saw at tublr a person who its making amazing [male] inflation videos using gen-2. But doesnt show up the prompt
>>2561 (OP)
Is there a way to download the character info from character.ai chats without an account? I wanna run the chat locally
Where do I find the most popular models/ safetensors for fat/ bbw for stable diffusion. Their used to be a thread for it, but I can't find it.
Any tips on how to avoid triggering content policies on DALL-E 3? Also, any ideas as to best practice in terms of the prompt format?
>>15445
Its trial and error unfortunately- start with [SAFE] at the start of the prompt. Then start simple and gradually add descriptors, seeing what ones seem to trigger the censor and try to work around that.

Whip out your thesaurus for fat terms. Add in "positive" words to balance out smutty stuff- toss in random, "proud, happy, confident," kind of words or describe bland things in the background- stuff that ideally wouldn't have any censor applied to it. The order in which you have your words in the prompt seems to matter too.

Then again, you might get a solid prompt and a run of good images and then you get censored a bunch with the same prompt. Thats the problem with Dall-E3- there is no rhyme or reason to it censoring stuff but when it works it can be pretty great.
Can anyone please tell me where to currently get the AuroraBBW model?
Does anyone have scripts or something to make spicychat bearable to navigate? Right click/Middle mouse don't work on 90% of the buttons, sorting by Alphabetical seems to just do whatever it wants instead, and that's not even getting into the search being about as functional as a child in a quartz mine.
>>15839
I also REALLY wanna know tbh.
How do I do progressive weight gain on stable diffusion? like making a set of images with the subject going from thin - fat
>>2894
I know this post is old, but is there even a slim chance of either a catbox or can you even share what model and Vae we're used here? This is soooo good.
How to generate good SSBBW pics with midjourney?
(76 KB, 827x662, stuffed_by_ovyaio_dgeuu8x-414w-2x.jpg) (56 KB, 827x662, stick_up_for_yourself_by_ovyaio_dgiim9w-414w-2x.jpg) (68 KB, 827x672, unwrapped_by_ovyaio_dgem7p7-414w-2x.jpg)
Does anyone know what models/loras are being used to generate images that look like oVyAlo/OIRTS/Ol Grizzle's images?

Here's some oVyAlo for reference. I used GAME and different character loras a while back but the images tended to look closer to anime and further from this more realistic/3d look, if that makes sense.

Some of Ovyalo's amazing work for reference
>>17351
That's his secret sauce. It's some custom model merge + hyperfusion + some manual labor.
>>17362

Yeah that tracks. I get that there's a lot of work going in behind the scenes. I'd just like to get a general base in this direction visually. Everything I do with GAME or other models I used a while ago is way more cartoony.
Does anyone know of any fat models, loras, etc that are compatible with SDXL?
(1.4 MB, 1024x1024, 15032-3825288353-Create a still frame of an extremely morbidly obese woman standing in a slightly messy apartment. She is extremely morbidly obes.png) (1.2 MB, 1024x1024, 15021-4202546017-Create a still frame of a young pretty woman lying in bed. She is extremely morbidly obese to the point where she can't move. Sh.png)
One issue I've had with SD is that a lot of models try to interpret the lower belly as a pussy. Is there any prompting tricks that can help with belly overhang without adding a pussy?
>>17527
1) simplify your prompts. Natural language can be useful at the beginning, but rather than giving SD instructions like you're talking to a graphic artist, think of it like tagging a photo for someone with vision issues. Just off the cuff, here's how I would have prompted that second image:

"an extremely obese woman stands in her home office, obese, huge belly, belly overhang, fat arms, brown hair, long hair, t-shirt, pajama bottoms, comic posters, computer, desk, messy room"

Once you have a pretty reasonable image, you can use a simple photo editor to pull out sections of the image and reprocess them with img2img if the fine details don't work. That'll give you a lot of control once the overall composition is how you like it.

2) add a weight-based LoRA to the mix. Symix, BGV5EX, Obese Girls, Chubby Girls just to name a few. Those LoRAs tend to 'understand' obese bodies better than any particular model will, especially if you're aiming for realism. Just keep the weights a little lower than normal (like 0.5-0.8) and it'll go a long way to keeping a large belly just a belly.

3) add "linea nigra" to your negatives. That's the vertical line pregnant women develop that a lot of the older models try to add to a large belly. They tend to 'think' that a really fat gut must be a pregnant person. It's not a magic bullet, but it's helped keep a lot of the larger bellies in check in my experience.

And I wrote a guide that details my process. It's not perfect, and I don't claim it's the only way to do this stuff, but you might find it useful.

https://www.deviantart.com/theguywhodidathing/art/How-did-you-DO-that-989204336
Dumbass question here: what exactly is "holo" in terms of style or art type? I see it referenced in the biggermerges readme where it comes up as "Holo vs Anime" with recommendations for different models.

Does it mean a more 3d style?
>>17681

Dude thank you for providing so much information about your process...I'm going to try implementing a lot of your recommendations today.
(33 KB, 512x768, do777j (1).jpeg)
I keep getting these weird creases up the middle of all my characters bellies. Is there any kind of negative prompt to make it stop doing this? I've played around trying to get it to stop for the longest time but no luck.
I've been trying out Dreamup AI through deviantart. While I've made some fat creations, does anybody have any tips on what terms I could type in to make them look closer to the 400-500 USSBBW range?
(1.5 MB, 1536x1024, Yotsu_31-12-2023_00027_.png)
Any tutorials anywhere on adding speech bubbles to pictures in general and recommended fonts? I can always prompt for a speech bubble but then i'm at the mercy of how big the AI decides to make it. I normally use GIMP for touching gens up after words but I dont see a module or template anywhere in there for speech bubbles unless i'm just really blind.
>>18540
I’ve just started dipping my toe into this stuff so take it fwiw, but I’ve had limited luck combining terms that mean “really fat” — “extremely fat heavyset morbidly obese” etc. That still gives random sized results. One prompt will give me a 300 lb woman, the next a plausible 450 pounder.
What prompt would I have to use if I wanted to make a woman/celeb look like she were swelling up with fat or inflating like a balloon or fully round and ripped/ripping out of her clothes? And what if I wanted to apply inflation to specific body parts?

Also, what AI would be best for that kind of thing?
(131 KB, 713x1121, are_you_really_going_out_to_the_club__in_that___by_youonlygetonebody_dg6ueho-pre.jpg) (95 KB, 669x1194, seed_oils_industry_lobbyist_advising_a_regulator_by_youonlygetonebody_dg9mx49-pre.jpg) (46 KB, 413x549, jeez_stacey__your_mom_is_getting_big___by_youonlygetonebody_dg1372p-414w.jpg) (43 KB, 414x517, you_ll_have_trouble_reaching_the_table_soon__honey_by_youonlygetonebody_dfw88ss-414w.jpg)
Faceapp did away with the "face swap" feature. Is there an app I can download that can replace that?
Hey, so I have about 60GB of fat arts from heaps of people on twitter, dA and pixiv that I've been scraping through 2023, for the purposes of making LORAs. However, I have no idea where to start, what settings to put, and am too retarded to really do this stuff without a step by step guide. Could anyone please point me in the right direction?

an example of one of my datasets: 150 pngs of varying dimensions in one artist's style, curated to only include blonde fat girls.
How many steps? tf do I do with the tagging? Looking at the training tab in the SD WebUI giving me an aneurysm.
>>7250
>>7251
kinda funny how using inpaint i have the exact opposite issue, i prefer bellies with flab and rolls and it mostly gives me big bloated ones
QUESTIONS:

-is ai capable to make inflation art?
-which site allows you to make your own fat art?
I had a pretty good run using bing for a while but recently the content filters were updated and my old prompts don't work any more. Even the stuff I could get previously was severely limited.

I have tried stuffer.ai and it's pretty bad. Light-years behind bing. I also set up a dalle3 account and tried generating from the CL but I've found the quality to be much worse than bing, despite both running the same backend.

What in you opinion is my next best option for AI generated SSBBW? TIA
(327 KB, 512x512, 00012-4207642774.png)
hello anons, I have been fucking around with stable diffusion, i managed to get something like this after a while. im going for realtively realistic immobility or near immobility.

you will probably tell me watch some tutorials but i'd rather just ask what you'd do as a next step and how you'd go about it.

say i wanted her belly to rest on the bed, i tried with the inpainting but that just kinda adjusts the edges but doesn't actually change anything big.
I wanna keep the belly button and make it look like there's cum dripping out which in this case was just a lucky gen.

have her arm and hand rest on her belly fold i also couldn't manage, it just removes her arm entirely.
do i hop into photoshop and change around things there and then have the ai fix it?

this is a bad case of just not wanting to read i must admit.
but if you like this i want to make more of it so help me out and teach me the craft please.
(466 KB, 848x512, BBW-Chan1.png) (2.3 MB, 1908x1152, BBW-Chan2.png) (2.8 MB, 1908x1152, BBW-Chan3.png)
>>20666
>>20729
I wrote a guide on my process which you can find here. It's not even a 20 minute read, and you can follow along as you're working on an image:
www.deviantart.com/theguywhodidathing/art/How-did-you-DO-that-989204336

The broad steps are as follows: Generate an image at a low resolution. Edit that image to get closer to what you want to see, then reprocess it in img2img multiple times, slowly increasing the resolution. Edit the resulting images to help guide the process, adding and removing LoRAs along the way. Touch up the final image by reprocessing specific parts like hands and faces, then edit the final result like you would a photograph.

In your specific case, here's what I did: The first image is just what happened after I ran your image through a blend of checkpoints (more info in the guide). Second image is what happened after I did my 'process'. I scaled the image up in PS a few times by 1.2x, each time reprocessing the image 4 times (at around .4-.48 denoising strength) and picking the best looking option. I only used the Symix LoRA at .5 strength the entire time.

Once I had it around 50% of target resolution, I scaled the image up by 2x in Photoshop, then selected areas that needed attention: face, hand, feet, and belly. When I reprocessed the belly, I painted a couple of grayish/yellowish marks around the navel and used the Cum On Any LoRA (https://civitai.com/models/24269) including the words Cum and Cum in Navel. I layered on each of the reprocessed parts, masked out the edges, and that's the second image.

This final image is after tweaking exposure, contrast, saturation, and skin tone. That's just photoediting though, and definitely not needed.

Hope this helps.
Ok so how can I turn my wife into a blueberry?

Obviously using her pics.

Basically put her in wonkas factory, or infront of an eating machine, like a kinda deep fake ai type thing if that makes sense?

Thanks in advance
Hi, wanned to start doing ai art and went to the guides thread, but some of the links there are outtaded or expired. Does anyone here has a up to date 2024 installation guide or video on how to do bbw art? Thanks
Hey all. Tried looking through this thread and searching through the catalog to no avail; not sure where else to look for a comprehensive explanation.

What is the best chatbot available currently, given the criteria of "good memory", "ease of character creation/customizability", and "not NSFW filtered"?

I have used C.ai since late 2022 but have become frustrated with things I can't accomplish due to either censoring, or me being able to draw things out *without* censoring and have them go along with it, only for them to forget the scenario we're currently in.

In a discussion with a friend last year, they told me Venus.ai let you do neat things that were explicit, but that its memory and response quality did not match c.ai.

Basically, I want NSFW chatbots, but not at the cost of overall decreased quality of conversations or interaction, and I am willing to pay for this.

What are my options?
Can somebody make a new gen furry thread im too retarded
does anyone else get a weird grayish hue to their images? It always seems to happen to me at the end stage of the creation process.
>>23949
250k cars for the gods lets piss off sergs
I've been thinking about a software that could be downloaded to import ai character jsons/cards into a software for some weight gain ai bots since I didn't really want to deal with the restrictions and paywalled stuff for some chatbot websites like Character.ai or CrushOnAI for example. If there is any softwares I could use that would be nice.
(114 KB, 828x1242, lara_croft_pitfall_2_by_scarebroart_dhii67p-414w-2x.jpg)
Anyone have a guide to making expressive faces? Or less that they use with SD 1.5, or prompts in PonyXL?

ScarebroArt pumps out pretty high quality pictures with fantastic facial expressions, and I'd love to be able to get anything close to these
What is the best model for immobile size so far?
Is there a way to prompt SDXL or more specifically PonyRealism to just make the belly bigger?

Because it makes the rest of the body fatter too.

Ok, i might have just answered it myself, the pregnancy tag. And I used it too at times...I need more sleep.
>>25542
Round belly, stuffed belly, big belly, bloated belly all will typically make the belly bigger without adding fat elsewhere. With symix anyways.

If you need to prompt pregnant for just a fat belly, make sure you add Linea nigra to the negatives to make it not look overtly pregnant.
(33 KB, 1026x790, 1e low angle.png) (159 KB, 512x512, 00103-1070069053.png)
can yall people who know what you're doing do something with this? its all about that perspective eventhough i fucked it already. how do i get correct perspective to work?
also generally what would be the next steps here to make something outta this?
Is there another ai program that allows you to generative fill a selection and use a prompt without any hard restriction? I used to use problembo but their new model sucks now
does ai understand prompts in regards to the "camera" we are viewing the image or video from? Like does it understand "walking towards camera" or "facing the camera"
(1.1 MB, 1152x864, 00019-1864446624.png) (1.1 MB, 1152x864, 00044-4116530323.png) (1.2 MB, 1152x864, 00043-4116530322.png)
Has anyone had success prompting for stuckage? I've tried a bunch of different approached (most recently in PonyRealism) but can never get it to work right.

I don't see much art of it either, so maybe there's no loras for it either. If anyone has had success then any pointers would be appreciated.

Here's some failed pictures that got close.
>>27108
Your "fails" are better than my best ever.
>>27127

PonyRealism is very forgiving, at least in my experience. Maybe it's just because I've messed around in Realism more, but I find it easier to prompt than PonyXL, at least in terms of getting what I want
(31 KB, 521x642, 1719941221110161.jpg)
>>2561 (OP)
Is there any AI/website that can do a reverse-prompt-generation for a given picture? i.e. giving artwork or photos as input and returning as many descriptors as possible in the form of a tag (bonus points if it includes some sort of percentage of likelihood).

i.e. giving a picture like this and expecting a list of tags in return (woman, bent_knees, midriff etc.)
>>27185

Yeah, I think AUTOMATIC11111 comes with a built-in Danbooru/CLIP reverser. Don't expect it to work miracles but it can give you a headstart
>>27108
try training a lora, maybe on civitai? you can also do img2img to gen some content if you don't have much source material.

Back to top