I have trained a base model from this dataset: https://civarchive.com/models/175874
An offshoot of the dataset of Bigger Girls Model with the addition of custom tags and a more curated (and re-tagged) selection of data. Intended to be a preview of styles, concepts and overall coherence available in future iterations of the Bigger Girls model, or at least in variants thereof.
Know anything about tagging images? Consider becoming a curator here: https://discord.gg/mw4nebTUu
Tagging
Underscores were removed as of V3. The legacy tags are at the bottom of the description.
The LoRA contains the following artists (more to be added in future versions), divided into high aesthetic and medium aesthetic:
high aesthetic:Akina, Aymusk, BeltBuster, BetterWithSalt, ccu, CoffeeSlice, DrWorm, Ekusupanshon, Fellatrix, Jaykuma, KipTeiTei, Lazorchef, ryokuchaMichi, Sizemologist, Squarewave, Tashoelle, xmasterdavid
medium aesthetic:BreadWithKetchup, MysteryDad, NewMetrack
The weight of the woman is controlled by the following tags, in order of weight:skinny, curvy, plump, fat, ssbbw, ussbbw
Hip size (as of v5) are controlled by the following tags:wide hips, gigantic hips, hyper hips
Ass size (as of v5.1) are controlled by the following tags:ass, huge ass, gigantic ass
There are also the following custom tags:gigantic belly, stuffed belly, hyper belly, full stomach, drooping belly, double bellystretchmarkstubefeeding, tubefeeding in mouth (usually needs sketch-inpainting to get the right hose orientation!)top heavy body, bottom heavy body, hourglass figurefeeding (WIP)stuck, stuck in doorway (WIP)burp (WIP)double chinbelly on table, belly on floorgurgling noise (WIP)clothed midsectionbackfatfupa
In addition to custom tags, the following "standard" danbooru tags are likely helpful:belly, big belly, hand(s) on own stomach, belly grab, thick thighs, wide hips, covered navel, midriff, midriff peek... (just search tags on danbooru!)
"Stretchmarks" guidance:
You may want to put "cuts, scar, scar on face, facial mark" in the negatives. Also, to put photorealistic, realistic in the tags if the results look too photo-real (I do have real pictures of stretchmarks in my dataset). And play with emphasis of "stretchmarks". If you use ADetailer for faces, put stretchmarks in the negatives (or do not include it in the positive prompt) for the face detailing.
Changelog
SyModel V3.0
New tagging system, principally that underscores were removed.
Many tags got another pass of revision in the dataset to ensure their validity.
SyModel 2.0
Converted to lyco. Custom tags should be working correctly again!
New styles: Fellatrix, Beltbuster
Fixed tag: ccu (instead of c-cu, which was broken)
Improved tags: cellulite, stretchmarks, fupa
New Tags: covered_navel (to complement covered navel vanilla tag), covered_belly, backfat
SyModel 1.0
Known issue: the custom tags don't seem to be recognized. This is fixed in the SyModel v2.0 version.
Same as 5.2 but trained as a model finetune and extracted to LoRA.
5.2
New training parameters used on 5.1 dataset. Same trigger words as before.
Reduced the size of the LoRA somewhat.
There may be more experiments in training in the future, rather than time spent increasing the dataset size.
5.1
new concept: gigantic_ass
Fixed tagging of ^^^ in dataset
New artists:aymusk, jaykuma, xmasterdavid
More data:squarewave
V5
Split hyper_hips into gigantic_hips and hyper_hips. There should now be more control over hips.
Artists added: ryokucha_michi
4.5
New artists:
high_aesthetic: squarewave (WIP)
medium_aesthetic: MysteryDad
4.4
Improved pregnant and outie navel content via increase in relevant data.
V4 (4.2)
Artists added:DrWorm, CoffeeSlice
Artists updated:c-cu (more content)
pregnant, outie navel --> focus on improving these. outie navels often don't look good yet (to be improved in a future version)hose, hose_in_mouth --> much better in txt2img. should make wrangling in img2img even easier than beforestretchmarks --> seems to have slightly improved. Faces can still be generated as "scratched" or "scarred"-looking. Can easily be fixed in inpainting (remove "stretchmarks" from prompt in face inpaint)reddish_belly --> still underperforms
V3 (3.1), First "stable" release
Replaced hyphens and spaces in custom tags with underscores, e.g. hyper-belly --> hyper_belly or pear-shaped body --> pear_shaped_body
more realistic stretchmark images in dataset
bodypaint should not longer result in tribal tattoos / tribal look by default
Trained on some of c.cu. About 1/3rd done with that dataset. Known as c-cu in tags.
Legacy Tagging
The LoRA contains the following artists (more to be added in future versions), divided into high_aesthetic and medium_aesthetic (which are tags):
high_aesthetic:Akina, Aymusk, BeltBuster, BetterWithSalt, ccu, CoffeeSlice, DrWorm, Ekusupanshon, Fellatrix, Jaykuma, KipTeiTei, Lazorchef, ryokucha_michi, Sizemologist, Squarewave, Tashoelle, xmasterdavid
medium_aesthetic:BreadWithKetchup, MysteryDad, NewMetrack
The weight of the woman is controlled by the following tags, in order of weight:skinny, curvy, plump, fat, immobile, fatblob
Hip size (as of v5) are controlled by the following tags:wide hips, gigantic_hips, hyper_hips
Ass size (as of v5.1) are controlled by the following tags:ass, huge ass, gigantic_ass
There are also the following custom tags:hyper_belly, stuffed_belly, belly_bigger_than_body (WIP, not reliable), reddish_belly (WIP), belly_overhang, double_bellystretchmarkshose, hose_in_mouth (WIP)apple_shaped_body, pear_shaped_body, hourglass_figurefeeding (WIP)stuck, stuck_in_doorway (WIP)burp (WIP)double_chinbelly_on_table, belly_on_groundgurgling_noise (WIP)covered_navel (WIP), covered_belly (WIP)backfatfupa
Description
All underscores in old custom tagging system replaced with space (" "), unless they are converted to something different as below:
artists:
ryokucha_michi -> ryokuchamichi
tags:
covered_belly --> clothed midsection
pear_shaped_body --> bottom heavy body
apple_shaped_body --> top heavy body
immobile --> ssbbw
fatblob --> ussbbw
belly overhang --> drooping belly
belly underhang --> drooping belly (tag unification!)
reddish belly --> full stomach, red belly (still a work in progress!)
hose --> tubefeeding
hose_in_mouth --> tubefeeding in mouth
hyper_belly --> gigantic belly
belly_bigger_than_body --> hyper belly
belly_on_ground --> belly on floor
under_belly --> belly over another (also WIP; not many instances in dataset so far)
FAQ
Comments (13)
theres a version on mega?
I absolutely adore this Lora. I've been using it since Preview 2 and it's my favorite "heavier" model. The custom tags are great.
Having given it a pretty good amount of time, I will say that 3.1 is my favorite. I love the shapes 4.4 gives (and they can go larger which is wonderful), but 4.4 doesn't play as nicely with Lora's and seems like it imparts some of its own style. It's tough to describe, maybe I'll try and generate some comparison pictures later, but 3.1 seems to give me very good results while leaving the model's style alone. Maybe it's just something wrong with my prompt? I don't typically like to use "Art by [artist]" tags because I like the Lora to sort of stand alone. But I was curious if that's you've run into as well.
Sincerely appreciate all your work on this. It's absolutely fantastic. Really looking forward to what else you make!
Add metalforever?
pryocynical
I think version 5.2 works really well and doesn't overide the style of any other Loras I use with it. I think the only thing it needs is more specific keywords to control the Belly, since it's hard for me to get it to do what I want.
I know I responded in a separate thread, but I just wanted to say in a top level comment that extract version 2 has been amazing. Awesome work!
What a fantastic X-mas gift! Thank you very much for your continuous work on the LoRa and the SyMix checkpoint, making both each time even better. The LoRa is really a masterpiece and always my go-to for getting more sexy curves, as it works really well both with 2D and semi-reealistic models. Keep up the great work!
this one doesn't seem to work on the in site generator, it either fails and never generates with nsfw turned off or with nsfw turn on generates but doesn't generate the Lycoris at all. i tried it elsewhere and am really impressed, only thing is it's a little hard to get certain poses with this one even though it does obese characters perfectly.
Which model did you use for the 3.0 version images ?
anyone else notice (btw, i'm a total n00b this is my second post/commet lmao) that no image generator is capable of producing an image of a ssbbw or fat woman in the same frame as a skinny woman or man? for example, i basically assume its impossible to have an image generated that has an SSBBW squashing a skinny man, everyone in the image is always fat... hopefully i'm wrong and about to be educated...
How was stuffed_belly and full_stomach separated? and can certain authors be tagged or only medium and high aesthetic? Great work so far!
See the PonyXL version here: https://civitai.com/models/397549
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.