For simulating pictures of people about to sneeze, obviously.
The formal training triggers are orgasm expression , screaming orgasm expression, and eyeroll expression , if you don't want rolling eyes, I recommend eyes closed, and wincing. mouth wide open is another useful prompt.
I biased the training pretty heavily towards adetailer usage, since this concept is mostly useful in conjunction with other LoRA's which tend to butcher the face anyway. Examples in the gallery. Have fun!
Description
FAQ
Comments (11)
Appreciated but the problem with this Lora and any LoRas is that they tend to override the effect of any other character Loras they are used in conjunction with. So with a known "celebrity" LoRa they will reduce the likeness to them. Not blaming you but just the way how things work with SD at the moment.
Yeah, you ain't wrong, although I do find it depends a ton a couple of factors.
The first is that my experience is that vast, vast majority of likeness LoRA's are unusable trash. Most barely look good even with the cherrypicked examples, and others will parrot a decent dataset-replicating portrait, but can't generalize about a person beyond that. I've only seen real generalizable success with LoRA's trained with method's like malcomrey's (dreambooth -> lycoris extract), which in my experience is capable of capturing things like body shape, skin tone, and subtle facial cues.
I also have found that only a few base models (like EpicPhotogasm Last Unicorn or Cyberrealistic 4.1 specifically, but I've also had some success with LazyMixRealAmateur4) aren't so overfit that they can then reliably generate well trained likeness LoRA's. Even newer revisions of those models have generally degraded their accuracy in my experience.
With dreambooth->lycoris trained LoRAs and one of the vanishingly small good base models, I have pretty good success in getting good expressions without too much style transfer. I find this especially with the prototypical expressions which require fewer training steps to encode, although when those fail it tends to be by not reproducing the expression at all. The more complex expressions (winking, lip biting, this one) which require more training steps all tend to result in more of a 'default face' bleeding through.
@thegipper
Thanks so much you your detailed reply and the links ❤️. I usually Kohya-train my own LoRas. EpicPhotogasm is my preferred base model.
On another issue of Openpose - it's disappointing that they don't just pose a figure but also change the body-type, eg. giving them big breasts etc. There must be additional info in them other than just pose/limb info. I find them virtually useless.
@FrozenGunner That's interesting regarding OpenPose. I haven't really used controlnet much lately, but playing around with it, you definitely seem to be right that the more of the steps that controlnet is included, the more it seems to revert heavily towards mostly rendering the model's "default" person. I think most of the default ways that the model knows to render a body come from he extra-horny data that was dumped in after SD15, so I bet there's a lot of bias in that data.
Also, interestingly I've found that counter-intuitively, the Tile controlnet mechanism somehow is more flexible with the subject that it's rendering than OpenPose, even with significant changes to the body type
@thegipper
"...extra-horny data" I'm sure that's the case. Lower the openpose value and you lose the pose... can't win LOL. I sometime set better results with DWPose but very randomly and often not much better than just prompting the damn thing.
This works great for me with inpainting. Just by tweaking denoising and lora weight I am able to get a good gradient for any level of pleasure I want without losing original style 😏 Thanks! just what I needed!
Curious - where do people use this type of Lora these days? I haven’t done any ai in like 5 months, but I used to use mage. That was the only one I knew that allowed nsfw - but no more. Does everyone use civit or local? Or is there another online platform?
I'm not sure, but I think a lot of people use their local machines with something like Automatic1111, or another SD web ui
i think most folks run SD locally and use automatic1111 or comfyUI. SD isn't censored if you run it from your PC. and jsyk you'll need a decent GPU with at least 8 gb VRAM, but more like 12-16+ gb. just google or search youtube for guides to install locally.
Great work! The most authentic orgasm-expression lora so far.
Can you please give an example on how you use it with adetailer? I.e. - Detector, Prompt, Inpaint blur, Denoise strength, Masked padding and anything else you use. All help much appreciated.
You can see the details by pulling the samples into Automatic1111's "PNG Info" tab, which will show you their generation metadata directly (including ADetailer configs). For that image I think most of my adetailer settings are just the defaults, but I copy/pasted the details below. That's actually a pretty high denoise, I more frequently use 0.45-0.55 for adetailer.
ADetailer model: face_yolov8n.pt, ADetailer prompt: "woman, screaming orgasm expression,<lora:oface_v1:1.0>", ADetailer negative prompt: "blurry, render, animated, cgi,", ADetailer confidence: 0.3, ADetailer mask only top k largest: 1, ADetailer dilate erode: 4, ADetailer mask blur: 4, ADetailer denoising strength: 0.65, ADetailer inpaint only masked: True, ADetailer inpaint padding: 32, ADetailer use inpaint width height: True, ADetailer inpaint width: 512, ADetailer inpaint height: 768, ADetailer use separate steps: True, ADetailer steps: 32, ADetailer use separate CFG scale: True, ADetailer CFG scale: 7.0, ADetailer use separate sampler: True, ADetailer sampler: DPM++ 2M SDE Karras
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.