A general thread for anyone to discuss bbw models, model training, model merging, or anything involving the mixing, or creation of models.

bbw-30k-step: https://mega.nz/folder/WFk13apa#WvCEO8IoqgSaeoKQILeIqg

LoRAs (contains multiple style loras for different artists such as pewbutt, kipteitei, and others): https://mega.nz/folder/sShj1LpS#Zh5ZxGO3uwK4QEaZ2X6z1w

Cocoa+Tea Model Merges and Chuuba Mixes based off of bbw-30k-step: https://mega.nz/folder/v80QhCCZ#TaTb_8vErCq0efZYdFh35Q

Some words of wisdom from /vtai/
Train your clip encoders!

>>2646 (OP)

Thanks for the Lora! BTW are they hosted anywhere besides MEGA? These would be excellent on Civitai or Huggingface

The kipteitei lora is awesome! I ran several of my OCs through it with minimal prompting and the results are phenomenal!

This make me want a KipTeiTei movie

>>2873
After this kip will delete himself, because he didn't have money from this

>>2873
But i hope that someone make it and we will replace kip forever, hehe/ We must replace all artists, their heresy will be banned

Does anyone know if there's a tool to rip images from rule34 with their tags for dataset collection?

>>2873
Woah, How'd you make this? Was it using an extension called prompt travel (or similar name) by any chance?

>>2888

Loopback wave. I made a thread here: https://bbw-chan.nl/bbwai/res/2826.html

>>2880

httrack?

Here's some additional resources for LORAs if you can't find what you're looking for on Civitai:
https://gitgud.io/gayshit/makesomefuckingporn#lora-list
https://rentry.org/hdglorarepo
https://rentry.org/trashcollects#loras (mostly furry)
https://rentry.org/gktwz

>>3915
When you get a lora that doesn't have a readme or any metadata visible in the Lora menu, is there a way to get the trigger words other than experimentation?

>>3939 Yes, there is. Lots and lots of practice and experience. It's the joy of learning to code. If you enjoy things like AI or trial and error then it's just the thing for you.

>>3939
Try to see if it includes any example images and take a look at those in the png viewer. Besides that, just guessing booru tags that match the theme of the lora

>>2871
What model do you use with the kipteitei LORA, I find your results amazing!

Is there a good guide on model merging beyond indiscriminate weighted sum interpolation?

I'm looking for a good model that just makes nice, small bellies. I've been using the bbw one attached in this but its not really giving me the results I want.

>>4000 No. That is why merging is not advised.

>>4095

I've had good luck with Starlike.
https://civitai.com/models/30373/starlike

I have some posts on the gallery page there. Should have my prompts embedded.

any reccommended models for realistic SSBBWs?

>>4158
There haven't been any model merges that have been focusing on realism yet.

>>4181
Could there be a Lora made for it?

Surprised no one has posted this one yet:

https://civitai.com/models/48197/obese-girls-or-concept?modelVersionId=52816

Probably the best resource I've found. Has solved most issues I've had with double navels, lapis lines, and the like.

>>2646 (OP)
Still pretty new to models in general, but Bigger Girls V3.5 just dropped.
Been using V2 for a few days, and still experimenting with stuff. Are the newer models generally better?
Link
https://civitai.com/models/2664?modelVersionId=58486

>>4277
I’m new to this stuff, so I’m not sure how to import it into Draw Things on macOS. I downloaded the files, could someone coach me through applying it? I tried importing it as a model, but nothing happens and trying to import it as a LoRA crashes my shit. Thanks for the help y’all

>>4393

I've had better luck running general models and using "chubby" or "love handles" in my prompts. The problem with Bigger Girls is that it removes data to get its results. The other ckpts have the same thick women that Bigger Girls has, BG just removes all the skinnies.

There's an ObeseGirls Lora that has some good success with anime. I've found it messes up more artistic models like Dalcefo painting but does well with Anything.

But really, just do a (chubby:1.5), (wide hips:1.3), (thick thighs:1.5) and adjust when needed.

>>4393
>>4277
>>4556
The merges are all much better than the stuff you can get out of LoRAs in almost all circumstances. (but still better than the NAI fats you get from general models)
Don't try and prompt wrangle a general purpose model into producing fats otherwise it will look like crap/NAI

If you are using base BG or bbw-30k-step you are doing it wrong.

Does anybody still have that Trinityfate lora? Link expired

Does anyone have a Butt Lift lora?

>>4578

When you say using base BG or 30k step is getting it wrong, I thought 30k step was generally usable (although the faces are kinda crap)

Are you merging 30k with AnythingV4, Abyssorange, or whatever other "normal" models you have? I know the BG model really needed a merge to be usable but im curious how people are using the 30k

Greetings, can someone make Lora on PreSiArt? I had only 4 gb vram so thank you someone who can make it.

Do anyone where I can get the bbw_hll3 basil-mixfp16 checkpoint

>>4578
The merges all barely work. By default they produce total blobs; the tag "plump" gives you huge girls.You have tag for every bodypart and crank the fatness way down to get something even close to approaching normal. They just give universally bad results.

>>4714

Its called ass support on Civitai

https://civitai.com/models/10373/asssupport

>>4875
>The merges all barely work.
That isn't true. All of the models merges work perfectly fine for both fats and non-fats.
>the tag "plump" gives you huge girls.
That is because you don't understand how to prompted with it. You just have to change your prompting style. The fat model merges, my example being GAME, will already try to give you fats without prompting for it. You don't need additional key words like fat, plump or chubby unless you want to go {{{bigger}}}. Prompt the body part being wide, or thick, and you'll get smaller results that are more what you are looking for.
Also, if you want really specific control, use the new aurora borealis model.
Pictures are examples that you can get non-blobs if you prompt correctly.

>>4752
I mean that the models are "concept" models, which in my mind mean that they have been trained in such a way that they sacrifice other aspects of their performance in order to ensure that the concept being trained for is done well. (Although this is a relatively new philosophy for me, bbw_hll 3.1b2 and bbw_hll 4 b2 are fixed up versions of bbw_30k_step)

They should be used as modifiers models, otherwise you need to merge a ton of models in order to make up for the deficiencies introduced by solely training the concept.

Basically it is harder to fix a deficient concept model than it is to add a concept model to a pre-existing model that works well.

>>4855
Look on the huggingface.

https://huggingface.co/mariaWitch/ExperimentalBiggerMerges/tree/main/Testing

>>4912
>You don't need additional key words like fat, plump or chubby

Congratulations, you now see why they are bad. If they're so limited that they can't produce normal girls, then what's the point?

>>4993
Except they literally can.
Just add fat to the negative prompt. ez pz
Like I said, I already changed methodology for bbw_Aurora Borealis. so if you want to prompt specific just use that.

>>4578

I rarely use merges on my stuff and I think they come out all right.

https://www.deviantart.com/fulforget85
https://civitai.com/user/fulforget85886?section=images

>>5010
The Aurora Merge is so versatile, that I think I'm not gonna go back to my previous models.
The ability to do semi-realistic, anime, pixelart, and low cfg images just makes it far and away better than any other model or LoRA combination.

>>4912
Aurora Borealis?

>>5019
Might just be me, but I'm finding it struggles with larger sizes.
BBW is does no problem, but it takes a lot of prompting and lora weight to get it to produce SSBBW sizes.
Tbf though, I'm probably doing something wrong

dumb question: how do you know when and which "clip skip" override setting to use? when using pnginfo on catboxed images from other anons, I often notice clipskip either being set to 1, 2, or not at all

>>5206
AFAIK, anything developed in 2022 against SD 1.5x was usually clip skip 1. So, older but still extremely popular and widely used models, usually in .ckpt format. Anything relatively recent within the past 4-5 months, is usually clip skip 2, and generally available in safetensors format. That said, that's just a rule of thumb and I'm sure there are exceptions. It should usually be specified on the models huggingface documentation or civitai blurb.

>>5207

Thanks for the info Barclay. That makes sense and I'll try sticking with Clip Skip 2 for now.

>>5010
I've been following your work for a while and I've noticed how you're able to successfully get characters eating. Most of the time when I do that, the food just sort of floats in front of their mouths. Do you have any suggestions for how to more accurately simulate eating?

Thank you for your help! I think your work is some of the best in this space that I've seen!

>>4912
>>5019
that's way, way too fat. I can't get any sizes that aren't total blobs out of these.

>>5746
>aurora borealis
>which is already known for having an overall smaller size compared to GAME
>is too big.
how to tell me you haven't used the model without telling me you haven't used the model.
Or better yet, post your images, so I can tell you what you are doing wrong, and help you get what you want out of the model.

>>5761
>says posted images are too big
>"UMMM ACTCHUALLY THEYRE NOT TOO BIG?!?!?!?!?!!11111"

ok. This is "1girl, plump." This is the minimum size Aurora is capable of going for fat girls. Everything you've posted is this big or bigger.

>>4641
seconding this, does anyone have a reup?

anyone tried merging anythingv5 with the bbw step model? i ain't exactly got the resources for it ever since collab died

bbw_Aurora Borealis Where can I find the model?

>>5931
https://huggingface.co/mariaWitch/ExperimentalBiggerMerges/tree/main

>>5779
remove 'plump' if you want smaller, try 'chubby' or 'bbw'?

the kipteitei LoRa is boomin

>>6003

straight bussin

>>5738

Use LyCORIS Swallowing Whole. Looks like it got booted off Civitai. Here, let me catbox it. It's a LyCORIS so you need the LyCORIS extension and put it into the LyCORIS folder.

https://files.catbox.moe/dc6x0x.zip

"swallowing whole" or "big belly" are some triggering tags. You can get the full list with the Civitai extension.

Check this image out for an example: https://civitai.com/images/723471

It does make the image overly red, so I'd use it at a 0.3 or 0.4 and then do a couple lower denoises without the tags in the prompt.

>>6016
>booted off civitai

why are they like this

>>6016
>and then do a couple lower denoises without the tags in the prompt.

wait, what do you mean by this?

>>6058

Exactly as it sound. Run img2img without the LyCORIS in the prompt at like a 0.4 or 0.5 two or three times. Helps remove the exaggeration.

>>6068
And that doesn't fuck it up? I always thought that would be like upscaling a jpg.

also that file isn't showing up in the Civitai helper, Would it be possible for you to upload your .info file too?

Is there a link for a google colab ?

>>6078
Just mount your drive and transfer it into the LoRa folder

>>6073

The file is here: https://files.catbox.moe/dc6x0x.zip

Doesn't fuck it up. As long as you don't overbake it. That means running a prompt for "cute girl" and then running a low denoise and high CFG for "dinosaur". When you overbake it, it looks over saturated and high contrast

>>5779
>This is the minimum size Aurora is capable of going for fat girls
Dunning Kruger at its finest. Learn how tags work and you'll be able to get what you want. This is with just [[plump]] and some prompts for the background

>>2646 (OP)
Is there a list of artists people use for their generations? Particularly with the Aurora model?

>>5779
This is literally a skill issue.
Prompt better.
Aurora is able to get very specific, the more specific you are in what you want, the more likely it is you get what you want

>>6210
I don't even use artist names when I gen with the Aurora model, good old prompting gets there for me.

If I was to want an artist style, I would use an artist LoRA

So what's everyone's opinion on Biggergirls v5?
https://civitai.com/models/2664?modelVersionId=85646
Able to hold a candle to the GOAT aurora yet in your opinion?

>>9024
These aren't even AI generated, wrong board

>>8953
Not nearly as good as GAME and Aurora, for vtuber content and anime stuff, they are still way better. I've been using rainbbw lately as well, top notch for fat girls and plays well with most size related loras

>>6210

would this be the aurora model by Luna on civitai?

>>9861
No.
The "Aurora" model people talk about here refers to the Aurora Borealis model here: >>5948

>>8953
BG5 while good, still has significant issues that prevent it from still being anything more than a concept model. (This isn't a bad thing though! Not every model needs to be good at genning, some may exist just to be merged.)

>>9105
Have you got a link to RainBBW? And GAME if you don't mind, can't seem to find either of those two

>>9979
Check in the thread man.
>>5948
And honestly I would stay away from plain jane fats models. Aurora Borealis and GAME are your best bet.

>>9901

Out of curiosity, what are the 'significant issues' that you're talking about that don't appear in the Aurora model?

I'm not really sure what you mean by it being a concept model versus it not being a concept model; is there something specific you're talking about?

I've also had a few issues with generating things with it, but I don't know that this isn't a personal prompting issue versus a flaw in the model.

>>10381
BG5 from the very nature of how it was finetuned has subpar backgrounds and texture details. GAME, and ESPECIALLY Aurora Borealis have much better texture, versatility, and backgrounds such that there really isn't any reason to use BG5, when Aurora does fats and outputs a better looking image.

>>10406
as someone who is looking to get a better idea of how to train a lora/model, what do you mean by how it was finetuned? Is there a better method out there?

Not disagreeing about other models, I'm curious about the best way to do it as someone looking to get into it.

Alright, big brain question time.
What is the best *non*-fat model in your guys opinion?
Was testing out the meinamix model and was amazed by how well it could do fat when using loras, pics related. These were all on the first gen too.

Made me wonder if using fatty models was the inferior method as opposed to using good models with fatty loras.
Discuss?

>>9979
I think this is GAME he's talking about
https://huggingface.co/mariaWitch/ExperimentalBiggerMerges/tree/main

anything in the works for sdxl yet?

is there an adjectivenouncombo lora?
if not are there any good tutorials for training loras without a booru?

>>10831

I like ReVAnimated. I usually don't use the BBW specific models since they're all anime based and seem kinda limited to that style. ReVAnimated can do anime but is pretty good at other styles too, I just use LORAs to make them fatter from a non fat base model.

>>13282

I have had great success just appending (plump,flat_chest:1.4) to my prompts. Granted you can probably drop the flat_chest as that's a more specific kink of mine.

My favorite model at the moment is Aniverse but I've had good success with Dreamshaper, Starlike, and epicRealism.

You can see all my stuff on Civitai - prompts included https://civitai.com/user/fulforget85886/images

>>2646 (OP)
Anyone got the better with salt lora? link to the folder is broken.

>>12316
I am working on a LoRA or a checkpoint, but I need a better dataset. I wish there were some archive enthusiasts who have a diverse collection that they shared on mega. I have found that I can get good SDXL results using the techniques outlined in the DALL-E 3 threads, but outside of that "morbidly obese" translates to "mid-sized". 30k images with diverse women/clothes(or naked)/backgrounds should get me there.

>>2646 (OP)
Anyone have a fixed link for the artist loras? I'd kill for that pewbutt model

What model is everyone using right now? Took a break for a few months and am behind what's new. I was previously using GAME and bbw_hll4b2.

>>17918
I typically use a non-fat 1.5-based model with SyMix. My favorites are revAnimated, CaramelApple, dreamshaper and MeinaHentai

>>17918
I use the obese girl concept lora and then just run my prompt through the 15 checkpoints I have until I get something I like.. Picrel is with aingdiffusion, but I usually get good results with sardonyxRedux or cartunafied. MixTapeBlues and Aniverse give good and consistent chubby gens but anything fatter doesnt work well.

>>17918

Last time I was experimenting with gens I was using models and loras based off of SD1.5.

I'm seeing a lot of discussion about PonyXL and other models, but there aren't as many lots for that one.

What are people using most commonly now? I went through Barclay's super helpful guide, but since it was written a while ago and I've fallen behind I imagine there are newer and better tools.

>>24902
Look like a troll lmao

>>24902
I recently made the big switch from SD 1.5 to PonyXL. Two things made that possible: the slow but steady conversion of many popular charecters to Pony-compatible Loras, and Symix being forked to work on Pony as well. This made all the difference and finally got me to switch. But yes, it's amazing that in a little over a year, that guide I wrote is basically completely out of date. I should do another one with all the latest stuff.

>>24904

Thanks Barclay. Your guide helped me get up and running initially, and I still find the concepts helpful. I've never used Symix so might switch to PonyXL and dabble. Looking to gen more photorealistic images than not.

>>24905
I'm also hesitant on switching to Pony as still it lacks behind regarding fat photorealism vs. many custom SD 1.5 merges and LoRas. But, already semi-realistic gens look pretty impressive plus more and more LoRas are available for Pony, so it may be a matter of short time...

Let's also see in a matter of week(s) how things will be once SD 3.0 is released.

Not sure if this fits the thread perfectly, but I've been working on a LoRa for SDXL (and PonyXL) base models.

Basically, you download this lora thing and throw it in the lora folder within models folder. Then you load up whatever base model you prefer and throw this one in the prompt along with a trigger tag of choice.

You can download it here: https://civitai.com/models/550134/refat?modelVersionId=612108

Feel free to give whatever feedback on it, I'll try to build on that later