This is for all the freaks out there. The aim of this lora is to help you generate realistic canine peens regardless of the overall image's style.
You'll see as part of this package there is a main "V2 Base" lora and a "V2 Detailer" lora. I'll explain more below but TLDR, using the base lora at 0.8 strength should satisfy most uses. Full strength is definitely possible but can struggle on more complex pics.
There are 4 penises in the dataset, named pp1, pp2, pp3 and pp4. pp1 is the "whitest" penis, then they get progressively redder until pp4 which is pure red. It's optional but adding one of those 4 tags can boost consistency with getting the type of penis you might like.
Now for the longer explanation. The base lora is trained on the 80 images of the 4 penises, with 20 context images added. The context images were manually created by me using some quick photoshop. They massively enhances the lora's ability to generate images with minimal style bleed and increased consistency compared to V1. The detailer lora on the other hand was trained without the context images. This means its really good at generating just a penis, which is perfect when used on your adetailer or fdetailer because that's all it needs to generate. The base lora 100% can be used as a detailer but the dedicated detailer lora is a little better at it so why not use it. I didn't use the detailer lora in any of the base lora sample pics but I definitely should've.
You'll notice I use a mix match of different methods and weights in the sample pics. Feel free to also play around and see what works for you.
For more info on how to use these loras. Here's a quick how to:
Generate your base, undetailed image the same way you usually would. Either add the base lora to this step or don't. If you do, add "c4nine pp" as a tag as well as any of the other tags on the right. You can stop right here, but if you want to go further...
Plug the detailer lora into either your adetailer or fdetailer.
Add the tag "c4nine pp" to your detailer prompt if it isn't in your base prompt.
Adjust the weight of the lora and the denoise of the detailer as you want. In my testing, a weight of 1 on the lora and a denoise of 0.40 is the sweet spot. The higher the denoise the higher the risk/reward.
If you're struggling, check the quick example workflows I have attached to the sample pics.
The dataset was generated using these 3d models. No real images were used.
Description
Trained on the below settings.
100 images:
{
"engine": "kohya",
"unetLR": 0.0005,
"clipSkip": 2,
"loraType": "lora",
"keepTokens": 1,
"networkDim": 16,
"numRepeats": 7,
"resolution": 1024,
"lrScheduler": "cosine",
"minSnrGamma": 5,
"noiseOffset": 0.1,
"targetSteps": 3500,
"enableBucket": true,
"networkAlpha": 8,
"optimizerType": "Prodigy",
"textEncoderLR": 0.00005,
"maxTrainEpochs": 20,
"shuffleCaption": true,
"trainBatchSize": 4,
"flipAugmentation": true,
"lrSchedulerNumCycles": 3
}
FAQ
Comments (9)
Thanks for all the support so far on these loras. I'm still learning but I feel I made a pretty significant leap with this V2. Big enough that its worth applying that knowledge to the monster penis ones. I'm a little burnt out on penises for now though so I might take a break for a bit but I'll definitely get around to it starting with the humanoid one.
Does version 2 have “larger devices,” or is it just the preview images that make it appear that way?
V2 definitely leans towards the larger side of things, but for the preview images I exaggerated it a bit to better show them off. If you look at the prompt of the images you can see I added "large penis" to a lot of them. I'm sure if you add the right tags and negatives you should be able to control the size in V2. Although tbh I haven't tested what the smaller ones look like.
@Boerboer ty
never ask a woman her age, a man his salary, and a lora creators where they found the reference images for their datasets..
I've definitely wondered how some people get there datasets before, but in this case I promise there's nothing dodgy. I used the 3d models linked at the bottom of the description. Plus my loras are trained on Civitai and they'd have a field day banning me if I used real pics.
Haha, honestly I'd still love to get a peek at other people's datasets... considering there are a bunch of pretty interesting LoRAs out there (plus their metadata is totally scrubbed). and btw, did you train this LoRA just on the 3D models from that link? I feel like that’s a pretty small dataset to build a full lora from...
@socksosock I go into it a bit in the description, but for this dataset I made 20 renders each of the different peens at different angles. Then I photoshopped them onto 2d drawings and made 20 images of that. So 100 total, with 80 being 3d renders and 20 being what are known as context images. Feel free to DM if you want the dataset.
@Boerboer nah don't worry about it lol, I'm just joking.
either way, the fact you rendered it yourself makes the lora way more valuable, great work
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.