Lying Down Breast-Morph / Backtits FLUX
Will not affect your other character LORAs facial features. New version trained on higher resolution images, making this lora far superior to previous versions.
Keywords: laying on her back, viewed from a top down view.
Suggested Weight: 0.9 - 1.0
Other Tips: Use together with the WEAK models on my Character Friendly Body Types LORA to better manipulate the body types.
Description
Better training data. should now not affect body shape at all, except for the breasts.
FAQ
Comments (8)
Can you explain your tricks used to train this to match the breasts of the character LoRA?
First of all FLUX is very smart and seems to be able to figure out much on its own. But i think the training data also matters, and you need to have a feeling what flux might be able to learn and not learn from the images you train it on.
When it works the results are great but I'm getting lots of body horror interspaced with the good generations, granted I'm using a few Loras together to get the tit/nipple shape I like, haven't really tried the lora on its own to see the result.
Yeah, i've done plenty of experimentation myself, and in previous versions i described it as something from the pits of SD3. I wish i could double check right now, but i'm pretty sure you'd get the same results if you deactivated this lora and still used a body shape centric one and tried to generate someone on their back. The LORA isnt perfect, but i would guess that it's more of a FLUX problem and the way it handles poses.
I'd assume that the LORA's you're using have been trained on upright bodies. And for some reason FLUX really wants to generate people upside down whenever you describe someone on their back. Maybe for some SFW reasons. So what happens is FLUX wants to generate a body with the head at the bottom, but then the LORA is trying to generate a body that has it's head at the top. Some prompts seem to work better than others however, so if you're getting more than a few of messed up results, change your prompt around.
I might be able to fix this by including upside down images in the training data, but since pretty much all other LORA's are trained on upright images i suspect it wouldnt do that much of a difference, and it would just double the amount of training data/time.
@mongot Yeah I think you have it exactly right in your comment, will play with prompting further..👍🏻
superb lora, now if we could get one for lower regions that didnt mess with characters lol.
i'll probably make something like this in the future. keep an eye out
this is so far the best that doesn't affect the face as promised. awesome.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.