Hi all, I've been encouraged to upload this embedding for wider use here. I feel it still has some shortcomings but perhaps they can be overlooked for now :)
This embedding is intended for creating POV deepthroat scenes and should do so with some success with the following caveats:
It is HIGHLY model-dependent - if the model knows this basic pose already it should do OK, otherwise it may fail completely. The images here were generated with the Ultra Realistic Porn Merge checkpoint.
Due to training data, it tends to create similar eyes and facial expressions. Not bad but a little restrictive
Still a little hit and miss in the best case, you may need to gen a dozen and img2img the best one to get a nice result
'Restore faces' can make the result worse
I prompt in the general format of 'blah, blah, a photo of a dpthrt woman, blah, blah', but it should be relatively flexible. Just reference the embedding name (dpthrt by default, but you can rename to anything) in your prompt, no additional trigger words required.
Description
46 images, captions, 5v, 46 batch, 858 steps varied rate
FAQ
Comments (6)
I wanted to understand how they get such a high batch, I have a 3060 with 12gb of vram and I can get a maximum batch size of 3 for textual inversion, is there something I'm missing?
4090 with cross attention optimization enabled. You should be able to get a much bigger batch as well if you turn that on.
about 12 gigs of vram TBH
Which model should I run this against? Uberporn? Can someone share an ideal prompt? Thank you.
same prompt, model + settings from example above does not give the same result. how is this supposed to work??
Why did you delete all pictures?

