Huggingface for lost Loras: https://huggingface.co/quadratrund
Please don't mind the frames pic in pic borders in the example images. I fucked this a bit up... the lora won't create this. Don't worry.
Found it on Illustrious and wanted to redo it. If you have other ideas feel free to write me. currently I am reproducing loras I like.
Should work on all image formats.
Strength: 0.7-1.0
Prompting:
against wall, head grab, cheek press, standing doggystyle, sex from behind, ass, girl is bent over, the photograph depicts an intimate scene between a woman and a man. the man, whose face is not visible, has light brown skin and is standing behind her, his one hand gripping her hip and the other hand is pressing against the girls head. he is pressing her head forcefully against a wall. he is nude and his face is out of frame, with his erect penis partially visible as he engages in sexual activity with her from behind. keep her outfit, surprised expression. Keep the environment the same.Description
FAQ
Comments (20)
I'm having issues with character diversity across seeds, it's nearly identical
I sense a pony majority in the dataset.
@OP: is that correct?
It would explain the face shift and the two left hands in the example image.
@Adaptalab0r yes this was one of my first flux Lora’s. For testing I used a lot of ai generated images. Currently reworking stuff. My Lora’s now use real world image datasets.
This struggles to maintain character identity.
you can always reduce strength and double load in the klora. also prompt: keep the girl, outfit and hairstyle the same.
@quadratrund This is the first I've heard of this technique, so load two Lora instances at ~ 0.50 strength in the same LoRA chain/powerloader? Or sandwiching a character lora?
@sonofabean Well I use draw things. I can choose the initial model and a refiner model that will take over at a certain point. When I loaded a lora I can choose that it should only be applied on both or just one of the models.
"Refiner model" is the term you want to look up.
@quadratrund I tried the half strength, and it helped slightly, but it looks like a base face in shining through. Maybe it's the tool I'm using, Wan2GP.
@Shaggy08 i'm using wan2gp and work with this strength
6 steps
step 1-2: 0.85 strength
step 3-4: 0.35 strength
step 5-6: 0.2 strength
😇
@zoefollonier can you explain please? In wan2gp what do you mean by the step vs strength breakdown.
@astronationz4740
I’m doing that specifically to preserve identity.
High strength in the first steps lets the action LoRA establish pose and motion, but then I quickly lower it so later steps don’t keep rewriting the face.
If strength stays high all the way through, the identity tends to drift
so i'm using 6 steps with "0.85,0.35,0.2"
@astronationz4740 I’m not sure if this is the right way to do it, but it works for me
@zoefollonier thanks! this made a huge difference. I forgot about the advanced lora controls in w2gp.
@zoefollonier oh, i didn't realize we can do this! thanks bro 😁
@zoefollonier Hey, you're absolutely right. I dug into it a bit and this indeed is the proper method to do this. Have you found a way for loras like SNOF to not alter the face much of the input image, or another general NSFW lora that doesn't alter the face much? Not through a character lora, just a input image. Thanks for your tip earlier!
@astronationz4740
Try version 1.3 of SNOF — it has better facial consistency. What works for me is lowering the LoRA strength a bit and adding prompts like: 'exact same face, same head angle, same expression, same facial features, same identity as the original photo.'
@zoefollonier Brother, I owe you drinks. Thank you!
@astronationz4740 I am not part of the conversation, but I like to see, how you guys help each other haha
@zoefollonier Yeaj snof better because it's quality over quantity
how could you do this strength increments in comfy ui?
Details
Files
FK_doggypressedtowall.safetensors
Mirrors
FK_doggypressedtowall.safetensors
FK_doggypressedtowall.safetensors
againstd-klein.safetensors
FK_doggypressedtowall.safetensors
FK_doggypressedtowall.safetensors
wall_dog.safetensors
Against Wall Doggystyle - Klein9B.safetensors
Against Wall Doggystyle - Klein9B.safetensors
klein9b_doggypressedtowall.safetensors
120-FK_doggypressedtowall.safetensors
137-FK_doggypressedtowall.safetensors
FK_doggypressedtowall.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.



