V2: Wetter and sloppier! Works well between 0.5 to 1.0 weight.
A lora that maybe no one but myself wanted. This lora has been trained to give that oiled up glisten to skin without being too wet or drippy. Base Hunyuan can already do oiled skin pretty well, though it often results in excessive drips and splashes; more wet than oiled. The oiled up skin also tends to disappear when using certain loras or especially when stacking multiple loras.
There's where this lora comes in. It works well on its own and appears to actually improve composition and overall look of generations. But it is ultimately intended to be used with other lora stacks to help bring out that glisten and glow you should see from a nicely oiled body.
Trained double blocks only at a high lr with minimal repeats. Trigger word is "oiled skin". Use with whatever weight you feel like.
I will likely retrain this with a larger data set as it tends to push toward lighter skin tones, likely due to the limited number of images used to train.
Use with the Hunyuan Goonmaker workflow to achieve similar results: https://civarchive.com/models/1315959?modelVersionId=1485540
Description
Doubled the dataset size and included some more diversity in skin-tone, poses, and squishy-ishy.
FAQ
Comments (13)
nice one. Just trained on body right? So shouldn't affect char?
I didn't blur faces, but majority of the source data focused on the body at various angles. So it might have some affect on character loras, though hopefully not a ton. Let me know if it does impact it more than it should and I might consider retraining with cropped/blurred faces.
just tested. works fine with a few character loras I have trained. Works well with a specific body lora I have too!
can't wait for this to work but it only seems to do busty women atm, is there a way to balance that out?
I have some examples with smaller chested women posted. Could it be another lora that's producing it? Or are you talking a want for flat chests? I recommend lowering weight if you feel the lora itself is driving toward busty women. It should still work in some capacity as low as 0.4 even. I will run some other tests too.
@olivereads38255 Did some testing. The issue of busy women falls with the Hunyuan base model it seems. Even prompting for "((flat chest))" or "((small breasts))" with no loras at all gives women with decent curves. You have to really prompt for it with something like "a petite skinny small nude woman with tiny ((flat chest))" for it to produce something akin to an a-cup. This can be overcome by tacking on some other loras that tend to produce skinnier/petite subjects based on their training data sources and lowering the oiled up lora weight a bit.
@KinkMaster hey you're spot on with this! thanks for testing it further, I'm finally getting some decent :D
Could anybody help me get a decent tutorial on hunhyan please? trying to understand the workflow. i hav only 8 gigs of vram.
Unfortunately, not sure how much you will be able to do with 8gig. Best to search here for hunyuan fast workflows or workflows for low gig. There are some decent workflows and articles to read through.
@KinkMaster I see. With how much vram are you training? using runpod?
@AlienPleasures To train decently you need at least 24gb vram. Runpod is an option. There is an article on here about training hunyuan with diffusion-pipe and it includes a 1-click runpod deployment.
@KinkMaster I see! Thanks!! Any links to the article?
Goodness me... This would be great with a "black tape project" lora. Also got me thinking bodypaint...