Rock hard, high profile, perfectly round, symmetrical, zero sag breast implants with small/medium size nipples. No teardrop shapes here. If that sounds good to you, you are in the right place. Note that this lora is focusing on shape rather than insane breast sizes, but the default size is still quite large.
I was extremely selective in the dataset to only include the highest quality Bolt-Ons of 99% pictures of real women only. (no sag, high profile, perfectly round and symmetrical, nipples on the top half of the breast etc.)
Version 2 should fix many of the issues from v1, like blurry eyes and nipples. Version 2 should also effect the face much less. Some people may still prefer the effect of v1, so i will leave it up.
Prompting:
For best results start your prompt with something like
A woman with b0lt0ns
A topless woman with b0lt0ns
A nude woman with b0lt0ns
etc. See the sample images and the prompts I used there.
I got best results with Euler sampler and Beta scheduler but YMMV
For an invoke ai compatible version see here:
https://civarchive.com/models/2447980
Description
Version 2 has been trained a a larger, more diverse dataset. There should be much less effect on the face and it is less over trained than v1. Some users may still prefer v1 because it effects the face more and works better with regular flux dev. However v2 fixes many issues i was having with v1, like blurry eyes and obvious overfitting. If you only have access to regular flux dev, and not a a realistic checkpoint like UltraReal Fine-Tune, you may also prefer v1.
Its still not perfect and i may make a v3 in the future.
FAQ
Comments (24)
Tried this lora, at different weights, and-or using different F1D models, it always generates a vertical blurry bar on the right of the image. If you look at the gallery images that people have posted, it's also visible in most cases.
Is there any trick to fix this? Or a v3 release which would address this?
I've noticed this a well and am working on figuring it out/what's causing it. Will respond here once I do.
I figured out how to fix the lines, there is a few specific blocks in the flux loras that seem to be causing them, at least for me, check this tutorial for comfy:
basically, keep the first block, but don't load the next three, my block weights look like this in the lora loader
1,0,0,0,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1
you can experiment with different values. in the next version (if i make it) i just wont train those blocks.
its not perfect but it helps get rid of most of them. I still see them sometimes, flux has a lines problem
@Cyb3rJunk good find, appreciate it. So basically, this is an issue with lora training that can either be fixed at the source (how a lora is trained) or during usage (the workaround in the reddit link).
Thanks!
@benzstation yep, you understand right. We are all still learning how to best train flux.
@Cyb3rJunk is a regular 4070 with 12gb (and 64gb system ram) enough for LORA training with flux?
@benzstation i dont think that enough VRAM, i do all my training on massed compute or runpod for all my training with around 96 gb vram and multiple GPUs
@Cyb3rJunk assumed so, that why i never looked into this. Appreciate the replies, thanks again.
i think this is the best lora but sometimes it give wierd vertical lines in the final result i tried lowering the clip strength to .6 and it reduced them but are still there. pls make an updated version of this if possible as this if the best i found so far.
see this thread for how to fix this for now. I will eventually make a new version. https://civitai.com/models/1106416?commentId=712185&dialog=commentThread
@Cyb3rJunk can you tell me which is the sampler and sechduler combo you use?
@bluedemon333 I mainly use euler, dpm++2m and deis samplers. I always use beta for scheduling
@Cyb3rJunk Thank you very much. If it is possible please fix it. Those boobs look really nice!
What are the silve things in the navel? How can I remove it?
Cant install on Invoke-AI 6.x - error - unknow lora type
same
I created a version that works in InvokeAI — you can find it here: https://civitai.com/models/2447980
For the best fake tits lora, now we need a z-image version!
Hey man! I’d really like to use your LoRa model, but after installing it in InvokeAI, it says it can’t recognize the model type. Both versions do the same thing. Could you do something about that?
I created a version that works in InvokeAI — you can find it here: https://civitai.com/models/2447980
@Mike_GodArt Hey thanks for making this, its not something i have time to do since i don't use Invoke. I'll add a link in my description
@Cyb3rJunk Thank you for cooperation! :)