interesting~

there is a bulge lora called bulge_slider, on civitai

ya, or turning testicles into dicks... we need control-net penis!

@Janet Actually, since you mentioned that... There is this, but it's only going to work if it can recognize that it's a penis in the first place.Some of the ones I've generated have been truly warped beyond my greatest nightmares. There's no way it would detect them as penises. But, in instances where the head is a little wonky or something like that, it should definitely help.

@m99 Yeah, that's the thing. Probably the best way to use it is to reduce the confidence threshold down to the point that it starts picking up false positives ("That's not a dick. That's a mushroom!"), then raise it again just a bit to keep those from happening. Then, inpaint whatever aDetailer doesn't catch.

But... there's another idea! The dick (and pussy) aDetailer model is an idea for a quick way to decensor artworks with bar censors and minimal mosaic censors. Just load the image in img2img with a denoise of 0, turn on aDetailer, and let it do its thing. I haven't tried it yet, but I'll do that.

@m99 That would probably work too and maybe better than my idea. I was thinking that most censored penises are still likely to be recognized as penises. I believe aDetailer should identify most of them even without training on censor bars and mosaics, and then inpaint them with better, non-censored ones. That is probably not the case with vaginas, however, which is probably why no one has tried it before now. Many if not most of those (especially the gaping ones) seem to be so censored that there's just a gray blur where it should be, so they would absolutely have to be trained on censored/uncensored pairs.

But that brings me to another idea: training an aDetailer model on pairs of deformed dicks and what they should actually look like, and let it go to town fixing all of the monstrosities. Wouldn't that be fantastic—body horror no more! If a model is trained specifically for dick errors, then it will recognize dick errors, and my previous comments about adjusting the confidence threshold and false positives wouldn't apply.

@dita I never tried aDetailer; it's just the same as inpaint but automatic, right? Because in that case, I'd rather just inpaint manually. Or is the qualiity somehow improved over inpainting?

@Janet @m99 Right. However, with a good dick LoRA, you'll get better results than not inpainting. And when you're making multiple images, it's much, much faster to have aDetailer do the inpainting for you. With the current version, you can use a prompt, sampler, CLIP skip, CFG, noise multiplier, and steps that are all different from your overall image. So there's really very little benefit to manual inpainting. The only thing you can't do with it is change the model. So if you want to use an inpainting-specific model, you'd have to manually inpaint. It even works together with ControlNet, although I haven't quite figured out how that works yet.

And like ControlNet, you can chain multiple different instances together. I ran a batch of 100 images the other day, with aDetailer turned on, using four instances. It automatically inpainted the dick, the nipples, the face, and the hands.

One thing I wish it would do that the older dDetailer extension does is to AND models together. With dDetailer, you can use a segm-based body detector and a bbox-based face detector and tell it to inpaint only the area where the two overlap. Which means it doesn't inpaint the extra space surronuding the face like it does when you use a bbox face detector only. Maybe aDetailer does do that; it has a "Merge mask mode", but I haven't tried it yet.

@m99 could you make an "eye size slider"? Ideally one that doesn't give googly eyes (the other one does). Just naturally larger eyes.

@m99 I'm not sure about the numbers. I've tried this

https://civitai.com/models/128553?modelVersionId=140768

which should be good, since Ostris is genius, but when I try it I only get googly eyes.

Other than that, I see a bunch of cartoon-eye loras. Maybe one of those work well with realism, but the ones I tried in the past didn't.

@m99 Hmm. That's unfortunate that you can't do that in Comfy. I've been using aDetailer quite a bit the last few days, and I'm really missing this ability to paint only the face. I may actually install dDetailer again and see if I can run it for face, and aDetailer for hands, penis, and nipples.

Incidentally, I tried using aDetailer to decensor an image like I mentioned. The only way it's going to work is to actually do a full img2img first. To keep from wasting time in img2img (and potentially changing the image in ways I don't want), I tried using a denoise value of 0 and a step of 1, so that it would more or less skip over that and go straight to aDetailer. But that creates a blurry image, about what you'd expect from any gen with step=1. I thought that maybe with denoise=0 it would do okay... But it didn't. So to use aDetailer to remove censoring, it's still going to take as long as it would to manually inpaint the image—or maybe longer. There is probably a standalone version of something like aDetailer out there, if I go look for it. That would probably work, but using it this way doesn't.

i'm guessing the reason dick loras do that is because the training data features them pointing both up and down, and the model doesn't understand that those positions are mutually exclusive and draws both at the same time

@spunkymcgoo That was also my hypothesis. But if that were the case, you'd think that someone would have made separate dick up and dick down loras by this point. And as for double scrota, I wonder if it might not be a combination of lumping in canine penises and calling the knot an additional scrotum; as well as images with men riding dildos, champagne room position, reverse suspended congress, etc., all of which would show four testicles, with two atop the others.