This is an all-in-one LoRA for various sizes of asses, thighs, bellies, and breasts. It aims to allow for generations of women with combination of just about any size of the before mentioned body parts.
Note that this LoRA depends IMMENSELY on canvas orientation and composition tags. For example, the gigantic tag may produce images closer to the huge tag if the photo composition puts the subject too close to the viewer (i.e. close-up or upper body). Larger size tags like hyper and colossal may not even be close unless the canvas is in square or landscape orientation. As a general rule of thumb, if the size you get is too small, try a wider orientation or try moving the subject away from the viewer. Also, larger sizes dump the face quality, so you will have to use a face fix if you use them.
A tagging guide is supplied to explain all tags that have been redefined or added and includes all tagging info. It acts as a general guide to what works and what doesn't, as well as provides a numerical value for each tag for those who want more info or to try pushing the LoRA.
There's also a workflow included now. It includes an inpainter to help address the poor initial generation of hands, feet, face, etc. when generating larger sizes. Feel free to check it out if you use ComfyUI.
To-Do List- Better face generation for large sizes to reduce need of face fix- Make the "self lying" tags operable
- Make the LoRA less reliant on canvas or composition
- Better define modifier tags- Better self lying consistency and posing- Make inflation work with more than belly tags
- Get all helper tags working- Fix mis-spelled tags
- Allow for "wide shot" type images with canvas shenanigans- Better tag recognition for colossal and hyper sizes
- Find a way to tag for multiple subjects- Get stuffed belly to stop generating stuffed animals
V2.0
Expanded the dataset a bit more and changed some training parameters. Most notably, the fetal movement and stuffed belly tags actually act as intended. Sizes reflect the training sizes almost all the time now. Some really under-represented tags such as hose in nipples even work now.
Furthermore, I redid my workflow for ComfyUI and attached it as a zip. It is commented and set up to help address the weaknesses of this LoRA.
I'm very happy with the LoRA overall right now. I'll continue to improve the little things like a few modifier tags that still are recognized and add more as time goes on.
V1.5
I pruned, retagged, and expanded the dataset. The gigantic and enormous tags have been slightly increased in size, the colossal and hyper tags produce better, more consistent images, and the faces produced along side the larger tags are at least salvageable.
A few new tags were added: wide hips, very wide hips, long breasts, very long breasts, elbow on own breast, elbows on own breasts, belly press, and sound effects. These are all helper tags to narrow in on the image you're trying to make.
A few other improvements were made as well. The fat and plump tags are more well defined and produce images with greater difference in body composition. Proportions (in terms of length of legs, torsos, etc.) are more consistent.
A few improvements and getting some stubborn tags such as fetal movement to work and this model will be ready for V2.
V1.4
I realized many of the problems I was having with this model could be remedied via tagging, so I went through the whole dataset a few times and added/modified some tags. I also corrected as many typos as possible.
V1.4 fixes most of the remain issues with this model. self lying tags are consistent across the board, size tags are better defined, belly tags finally work as intended, and almost all included helper tags do what they're supposed to.
I have a few things to finish up, like getting the fetal movement and no navel tag to work. I won't release the next version until it's well enough along to be considered V2, and I'll start support more base models after that.
V1.3
Another slight improvement to the model. The stuffed belly tag can actually generate the desired shape, and all sizes are a little more consistent. Some tag combinations and views are still slightly off, and the positions associated with the self lying tags are limited. The next update will focus on the self lying tags and any other tags that are present in the dataset but don't work (i.e. round breasts, fetal movement, all hyper tags together, etc.).
V1.2
Overall improvement to the model. better defined sizes, less tag weight to get larger sizes, helper tags having more effect, more poses are recognized, etc. The only drawback I've noticed is that the dataset is lacking behind in bottom heavy images since I've been incredibly picky about those. You may have to ad a little extra weight to ass and thigh tags when they're paired with belly or breast tags.
On another note, some of these images probably have the wrong prompt associated with them. The metadata is correct, but it may be displayed as something completely different on Civit. This is just the result of growing pains learning to fully use ComfyUI.
V1.1
Expanded the dataset by another 1000 images or so. The larger tags generate better, the self lying tag actually generates the desired result (most of the time), and stuffed and pregnant bellies are better defined. I'll be adding and improving modifying tags such as belly shapes or breasts shapes with the next update.
V1.0
After manually tagging ~3300 images, failing, and then manually tagging ~3800 images, I've finally managed a working version of this LoRA. Currently, the only consistent generations can be accomplished with the huge-colossal sizes. Hyper will generate, but it isn't as refined as I'd like it to be. Other modifier tags, such as breast or belly shapes are weak as well. I also tried adding a tag for when the subject is lifted off the ground by their stomach or breasts. but have not managed any generation that actually depicts this. These issues should be remedied by the next version
Description
Stuffed belly tags works without adding stuffed animals
More consistent self lying
Added small calves and long legs as helper tags
FAQ
Comments (27)
Honestly, I've been struggling to get no navels consistenly for years by this point, if you can actually get the no navel tag working consistently you'll be an absolute legend in my books considering how seemingly every single artist that does large pregs exaggerate the navels.
I'll certainly try. I found I had to add more than one helper tag to get certain belly shapes right, so I might try that with no navels and see what happens.
I would love to see you add spherical or similar sized stuff, given the hyper tags already, or add blueberry style stuff. Then you'd have a lora with everything possible in it.
I'm considering it. I plan on getting just the inflation tag working first to see how hard it is to get the LoRA to differentiate between expansion and inflation. If that doesn't prove too challenging, I'll try adding it.
This is peak
Any news on V2 or has life just gotten in the way?
It's on the backburner, but still planned. I'd like to have it done in the next month or so.
Do you have any advice on how to move the subject away from the 'camera'? I've been struggling to do so, as the model doesn't seem to want to go very far away with image composition tags alone. I've had reasonable results with wide shot but that's about as far back as I can move her.
The only thing I could think of is purposely making the canvas wider than needed for something like a wide shot or making the weight of the composition tags rather high, but I haven't much success with that myself. If you're using a workflow that allows you to adjust the LoRA model weight and clip separately, try setting the weight of the LoRA around 0.6 but keep the weight of the clip around 0.8. No idea if that'll work but it's the best thing I can think of. I sometimes use this LoRA to help with that if it's an indoors shot:
https://civitai.com/models/1309301/modifier-larger-rooms
I'll definitely call out in the version description when I get the LoRA good enough to do this at some point in the future.
@reggiemccorthy1457 Appreciate it! This model makes big hyper SO much better, I'm so glad it exists.
Oh and worth noting, my frontend doesn't support clipping on SDXL derivatives, so that's something I unfortunately can't mess with.
Really nice lora you've made so far, looking forward to when or if you manage to get a smaller girl / zoom out option.
Any chance you would think about adding hyper/ giga futa support in the future?
Thanks! I don't have any intentions of adding futa stuff to this LoRA as of now. It's not really my cup of tea and the amount of data I'd have to add would be a lot. That and I don't even know where to begin when it comes to tagging for that.
An issue I've been having is interference from character LoRAs. Because those kinds of LoRAs train on general pool data and typically require pretty high strength to work well, a lot of times their definitions of "huge" and "gigantic" breasts fight with this LoRA. Turning up Hyper's strength to compensate works to an extent, but also starts impacting the effectiveness of the character LoRA.
Is this anything fixable, or is it just the nature of the beast?
This is something I have noticed as well, the generation appears to be going fine up and until the past part where, abruptly it seems, at least Hyperv1-4 smudges the face and leaves artifacts so bad that Adetailer's base settings can't identify a face or mistakes the entire breast for hands, or even faves. Most of the gens in the page seem to use civitai gen, I am using local gen A1111.
It really does clash with character Loras, but seems to work fine woth style Loras. I'm still prompt tweaking as we speak and the only method I have had that had success were character loras that have alot of head and face details, but its not guaranteed to work.
Sorry to piggyback, but I do hope a solution shows itself, whether through community comments or an update to the lora.
@cowenthusiast By all means piggyback, keeping the topic consolidated is probably helpful for OP to review.
My personal face workflow involves a lot of inpainting and regionally restricted re-generation so I don't have a problem with faces being bad. I usually fix the eyes via a LoRA anyway. It's just hard for me to get sizes right when a character model is "trying" to shrink the breasts.
I think this is a product of the exact tagging methodology for that part specifically and that's probably the source of that conflict. ("huge" and "gigantic" breast tags are available in the common image pool and character LoRAs would be using that size metric)
I'd assume style LoRAs (which I don't use, doesn't feel right to imitate a specific artist's style) conflict less because the target artist probably does more content like you want to be generating and so their pool of data is skewed better in the direction of the Hyper LoRA.
@Firesoul10 Would you happen to have suggestions as for inpainting guides, whether through adetailer or img2img? I'm still very new to using those tools; As if i don't get it in one with text2img, while i have very good base images, save a few artifacts and hand problems, the heads/faces will always skew and seemingly mess up last second in the gen. I'm just tired of only being able to fix it when adetailer catches that the color that was left smudged is a face.
Although, when its not giving the girl blackened hands/arms/legs/feet, the new Deeper Skin Lora slots in perfectly: https://civitai.com/models/1701939/deeper-skin
I cant really means test as i'm running on very dated hardware, I am making do with the gens. that said, 'wide shot,' and 'very wide shot,' help in preserving the size.
@cowenthusiast Those tags do help, yeah, but it doesn't really fix the interaction with character LoRAs. I can't give an in-depth guide right now, and I've never used this site's frontend so idk what it does.
The keys I've found for inpainting are:
Shrinking the generated region. The software I use, Invoke, lets you move the bounding box on the canvas, so I shrink it while inpainting as small as I can get it without losing relevant context. This means that while re-generating the face, I try to include the whole head so it knows what it's looking at. A smaller region of generation means it both generates faster (because you're asking it so make a smaller image) and allows it to not have to think as much about details besides what you want to fix.
Adjusting denoising strength. It's a value that controls how much of the original generation the AI is allowed to change. I've found that 0.3-0.4 is good for minor blending, 0.7-0.75 is good for minor detail fixes, 0.8 is a good general strength is the level of change desired is moderate, 0.85-0.95 is if you need a MAJOR change like a recolor or a re-pose, and 1 is the nuclear option.
Hope this helps, though idk how much of this applies outside of the frontend I use.
@cowenthusiast (btw I'm ENTIRELY self-taught so I know about as much about guide-finding as you do lmao)
@Firesoul10
As of right now, that's just the nature of the beast. Having beat myself over the head trying to figure out how LoRAs work without a proper education on it, it seems combining two LoRAs with wildly different CLIPs will water down one or both of them. I think the only way to fix it is for me to learn more about how training works on the neural level so I can make my LoRAs a better generalization of the data instead of being too heavily trained on it.
Also in regards to face inpainting as I have to do the same myself on many images: I assume it's given that SDXL is set up to generate in 1024x1024 variants while the images for the larger sizes are usually a way bigger canvas than that so the subject can fit and the face can still be detailed. I haven't dared to test my theory since I I don't have the VRAM. I would assume that starting with a larger latent would help, but I'm certain that would bring about it's own set of problems even if it helped the face.
I'm testing a newer version right now that'll hopefully iron out some of the limitations of the older models.
@reggiemccorthy1457 Yeah, I feared as much, given my understanding of how these things work. Thanks anyway!
Another thing I've noticed is that the LoRA has a strong bias towards belly when mixing parts. Don't get me wrong, I love belly, but I've also been trying to get a very topheavy but overweight build to work and it's really unreliable. The slightest hint of belly tags seems to cause it to overcompensate. I've tried top heavy but it doesn't seem to do anything.
It's probably some combination of belly tags trending much more extreme in the dataset and the AI trying to demonstrate tag compliance too hard.
I picked up on few things like that. Tried a v1.5 out and it was even worse about it, so I've decided to completely retag the dataset. I think I include the top/belly/bottom heavy tags too often so they ended up not having the desired effect.
None of ur updates r active on Pony Diffusion
I'm going to release another Pony model once I get the LoRA to a point it's doing most of what I want it to do. Training this model takes way longer than the others I've made and it gets pretty tedious to train for more than one base model each time I try to improve it. Hopefully the next version will be good enough to consider doing this.
Any news on V2 or has life just gotten in the way again?
Couldn't have picked a better time to ask. I'm currently stress testing a new version that I will be posting as long as all goes well.