So I've found SD is excellent for 'solo' / 'pinup' / 'softcore' / 'portrait' / 'sfw non-interactive' type shots out of the box, and pretty darned good for 'hardcore' / 'nsfw' / 'smut' also depending on what model you use, how you prompt, whether you use ControlNet, and whether you're an noob / idiot like me, or a seasoned pro. It also helps if you don't mind going through a ton of stray appendages, misshapen body parts, and 'nightmare fuel' glitches aplenty before you get it right on the 50th tweak to your settings :P
What SD is not so good at ? Kisses. The simple romantic gesture / interraction which involves the intersection of two human photgraphic subjects, overlapping facedata, and a mass of confusion for the algos no doubt, no matter what model, sampler and combination of settings you use.
So much like 'Legshow', here's a Lora / Dreambooth to try to solve this perceived problem.
Trained on a small set of 42 photos for v1. We've got a mixture of 100% sfw and some nsfw in the training data (so I've found bare boobies and/or chests creep in depending on strength and prompts), we've got boy-girl and girl-girl (although it seems to work for boy-boy too as you'll see in one of the sample images I generated once trained) and I've even included a small 'easter egg' with a completely NSFW variation (haven't tested that but the rest works so I'm sure it will too) for a few training shots too.
The 'magic woid' is "properkissing". "French" may work as a modifier, so may "with tongues out". Strengths are really random from what I can tell, based on a combination of model, subjects, other embedding or LORAs, the will of the gods ... who knows :-D
Let me know how you go, and please DO share your outputs as I love to see 'em !
Description
FAQ
Comments (18)
Why are your images tagged 17+ ?
I have no idea - that's Civitai, not me
Love!
Damn, the LoRA is very horny, hahaha
I know ... may dial that down a bit in further training. Tested my 'easter egg' too now and ... daaayum ... works even better than the 'main' Lora I think. Think properkissing the 'tip' of something if you want to give it a whirl 😉
@AutomaticDesire Oh, I see!
@XX007 may end up 'splitting' versions - Properkissing-SFW + Properkissing-NSFW ... what do you think ? Obviously more training and refinement for both if I do that ...
@AutomaticDesire Just gave a try to the easter egg...oh you! As for the splitting, I'm not familiar enough with training to tell if it would be more work to keep one model and make it more flexible or splitting it in two and trying to keep them strict.
@XX007 oh I'm still an absolute noob, learning as I go. I'm sure there's a more efficient way to do it, but I'm just planning to train again while varying my training images. The question is just whether I keep the one Lora and add more of both nsfw and sfw to it, or split it into two Loras and keep SFW training images in one, and NSFW with the other. I'm leaning towards the latter, since it means people can choose SFW or NSFW without having to do extra work in their prompts to encourage one or the other, like they do if the one Lora includes both
For some reason when I try to load this through the built in Lora extension I am getting errors, but through the extension works. I would prefer the built in version like in your examples, but I am not sure what the problem is... Any thoughts?
activating extra network lora with arguments [<modules.extra_networks.ExtraNetworkParams object at 0x0000026345FA28F0>]: RuntimeError
Traceback (most recent call last):
File "C:\stabledif\stable-diffusion-webui-git\stable-diffusion-webui\modules\extra_networks.py", line 75, in activate
extra_network.activate(p, extra_network_args)
File "C:\stabledif\stable-diffusion-webui-git\stable-diffusion-webui\extensions-builtin\Lora\extra_networks_lora.py", line 23, in activate
lora.load_loras(names, multipliers)
File "C:\stabledif\stable-diffusion-webui-git\stable-diffusion-webui\extensions-builtin\Lora\lora.py", line 214, in load_loras
lora = load_lora(name, lora_on_disk.filename)
File "C:\stabledif\stable-diffusion-webui-git\stable-diffusion-webui\extensions-builtin\Lora\lora.py", line 176, in load_lora
module.weight.copy_(weight)
RuntimeError: output with shape [32, 320, 1, 1] doesn't match the broadcast shape [32, 320, 3, 3]
I'm sorry @robproctor83 - I've literally only been using SD (AT1111) for 5 or 6 weeks (tops) myself at this point :( Magespace for about 3 days before that (paid for a month of premium, did not renew), before I went "fuck you, I have a 'gaming' PC I bought with my 'golden handshake' from my last job, GPU is decent even if not 'top of the range' .... let's investigate this more" 👍
I'm an English major ... yes I've been online one way or another since 1991, yes I followed up the basic college degree with a post-grad in multimedia (CD-ROMs and web design) before 'dot.com crash' #1 happened, and a 'vocational' course in 'hardcore' I.T. ('Business Systems') after that. I've also worked in online / digital media for the last 16+ years and learned + forgotten more coding than the 'average' Joe learns in a lifetime. SADLY however when it comes to Python + AI algos + anything beyond a very basic understanding of how SD actually works ... you and I are PROBABLY on the 'same level', and it's quiet possible you know MORE than I do. I hope someone in the community can help you, brother. I'm sad (but honest enough) to say I have NO friggin idea. Sorry - good luck !
@AutomaticDesire Thank you for your kind and honest answer. I have looked into the problem a bit and it seems that it was related to a recent commit (22bcc7be) I updated to. It may already be patched in the latest commit, I didn't check, but for anyone else who may experience this you can update your webui built-in lora with these 2 edits to the py file
@robproctor83 Oh my gosh, I was getting the same error because I had the same commit, and the same fix worked for me too. (Although for me, the Lora wasn't working.) Thanks so much for digging up that solution!
@AutomaticDesire And don't you dare apologize. Trying to use anime kissing loras for real people, uh, doesn't work well, but this one does! Keep at it!
This works for me as a Lycoris not a Lora.
Thanks! Can I use this with LatentCouple to make two LoRA characters kiss?
I don't know, not familiar with 'LatentCouple' :) Give a try ?
For a really odd and strange reason I get a message saying "Failed to match keys when loading Lora C:\Diffusion\stable-diffusion-webui\models\Lora\properkissing_v1.safetensors:"
And then followed by a huge brick of error text.
I really wish I could use this LoRA, not to mention that practically no other LoRA works for me. I've no idea where to even begin to troubleshoot this problem!
I'm running Stable Diffusion on my CPU, if that's any hint.
I got that too. The reason is because this is really a lycoris not a lora so it goes in models/Lycoris not models/Lora, and if you don't already have it then you need the lycoris extension which is in the list of available extensions called a1111-sd-webui-lycoris.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.