Republished as someone literally paid for this to be used in Civitai's generator :doro:
Concept LoRA based on "Venus Body" type, this need no presentation if you have used mnemosynekotorin's LoRA for SD1.5
My version of this LoRA seems to work fine on most anime SDXL models, this was trained on 1052 images and aims to produce thick thighs and wide hips even if you don't increase breast size.
This has a trained Text Encoder, please remember to test it as well, all preview images were with TE strength 1.
if you have a dataset that might increase this LoRAs "effectiveness" please do tell me how to reach you.
Caveats/quirks/issues
Has a trained Text Encoder so I truly don't know how this will behave with some models that seem to have a trained TE (For example it works with Animemagine XL but has a ton of issues)
Known trigger words for chest size, "flat chest" "big breasts" "huge breasts" "gigantic breasts"
Known trigger words for hips and thighs size, "small hips" "small thighs" "wide hips" "thick thighs" "huge hips" "huge thighs" "gigantic thighs" "gigantic hips"
Increasing breast sizes will increase the size for thighs and hips, sorry.
Using "slim waist" "slender waist" "small waist" will reduce the size of the waist, hips and breasts, please experiment with it.
Conversely, "plump" and "plump belly" will increase all 3 sizes as well.
Can generate kinda muscular looking females, arms, thighs, abs, and back mostly
KINDA works with DreamShaperXL, but haven't tested on other realistic models, sorry.
Most of the previews are made with ReproductionXL v3.1, again, sorry.
Used SDXL v1.0 with v0.9 VAE as base for this model.
Since SDXL has some issues with NSFW I tried training without much NSFW content. That's why you'll see areola, areola slip, and nipples in the NEGATIVES, since the dataset used for training had images with those, and putting these in negatives seems to reduce some sudden nudity creeping out of nowhere.
For whatever reason using wide hips and thick thighs will sometimes, produce a image like it had "from behind" tag, dunno why, but seems to happen with these tags, using a "facing forward, facing front" or the like, solves it, again, dunno why.
Expressions are mild but that's due to the lack of 'em in the base models.
These were the custom nodes installed in my ComfyUI installation
This is a bit outdated but I've included Adetailer for faces in the full workflow included with the images
Has two sets of upscale workflows just connect the node you want to run, both uspcale with model, and latent upscale + model, you can connect both to a Adetailer for the face, the full workflow is ready
If you like what I do please consider supporting me via Ko-fi, training SDXL is insanely expensive both hardware-wise and time-wise.
Description
v0.9 Alpha, heavily trained on Anime
FAQ
Comments (16)
For some reason this Lora doesnt show up in my webui, the list shows every other Lora, only this one is missing.
To be completely honest I used SDXL with ComfyUI and didn't have any issues there, In A1111 you might have to put it in the LyCORIS folder since white this is a LoRA it has convolution info.
@bluvoll There is no LyCORIS folder, LyCORIS files are loaded from the LoRA folder now.
@leolawliet777532 Then I honestly don't know how to help you, since ppl have used and reviewed it with A1111, I'm truly sorry.
having the same issue
@bqh70144 I tried using the lora with Kohaku XL and didn't had any issues, but I was forced to load the model before trying the LoRA, dunno if that helps you, sorry
This is an sdxl lora, meaning you need to use it with an sdxl model. If you have a non sdxl model loaded, sdxl loras will not show in the webui and vice versa; if you load the sdxl model, loras made using sd 1.5, for example, will not show.
have you any plan to use animegine3 for trainning for venusbody lora?
Not in the near future, SDXL is still a pain to me with concepts and styles, plus even a 3090 takes a ton of time just for testing and some really really bad results, I currently lack the motivation and the money to try to repeat this lucky result with Animagine, sorry.
@bluvoll If so, could you please provide me with the dataset? These days, I am training XL Lora using Colab as a hobby. Even if I make it, I won't share it anywhere else and will just test it personally.
We need the official to release!
This kinda works with Animagine but has issues, as I said in the other comment, I currently lack the motivation and the money to try to repeat this lucky result with Animagine, as my energy costs are 0.2 USD per kWh, and just training this lora it took me some solid 36 to 40 hours.
So this Lora appears to work when I use ComfyUI but not when I use Automatic1111
this is old, this is alpha but still the best one!!
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.