Dont waste your buzz, I dont do civit stuff.
Just sharing my own personal loras Ive been experimenting with.
Nsfw Upscale
Edit lora: 1 input image
Specifically for high frequency details.
Trained on high quality nsfw explicit images.
Trigger: enhance
Edit: workflow is outdated, will update it soon, check comment by other user below
Required:
Add the LTXVpreprocess node, it adds a predictable artifact to the image, that this lora was trained on to fix (native comfy node). ImageScaleToTotalPixels to --> 2MP.
There should be a workflow embedded in the image. Pretty straightforward though
Nsfw Transfer Head
Edit lora: 2 input images
V1
A somewhat better head swap lora for nsfw.
Female centric.
Trained on various viewing angles with some facial expressions.
But it seems that klein just struggles with understanding erotic expression in general soo.. dont expect too much.
Trigger: transfer head
Full caption it was trained on: transfer head into the same position and location with the correct size while retaining the original emotion and expression
V1.1
Removed some junk from dataset and retrained with better params
Trigger: transfer head
Actually found that mentioning expression or emotion is bad because it triggers the base model to override the lora with dumb shit
Nsfw Forbidden Lora
Edit lora: 1 input image
Its a lokr. Old one from when I was playing with lokr's.
Female centric.
Trigger is in the caption details.
Description
FAQ
Comments (4)
usual worklfow with a minimal change is much better .. for nsfw use any lora is close to content
Really good! I have been testing upscaling/enhancing especially lowquality nsfw photos for a while now, your Lora is a pretty good consistency keeper while adding subtle detail. Before that I used a consistency + SNOF 1.3 Lora. Yours delivers sometimes better results!
The Lora seems to struggle with body hair, but its a common problem with klein I think.
I have been using strength .85, what would you recommend?
For v2 I'll add a few more pics of body hair and maybe do a higher rank, hopefully that will help.
0.8-1.0 depending on the image
Thanks for the workaround but hopefully we get a patch to the model itself to fix upscaling saturation (and contrast!) shift. At lower quants kleins really blow-out the detail.
