<h3 id="heading-21">About V3.0</h3>Updated dataset and used multires technique in this version. It would bring more light and shadow with quality tag, such as masterpiece, highly_detailed, best_quality ,etc.Bewared that may affect the result of generation. You may get chaos result if you use these tags and overweight them.Multires technique parameter:multires_noise_iterations="6"multires_noise_discount=0.3<h3 id="heading-52">Introduction</h3>This LyCORIS is used to generate futanari with hyper penis.<h3 id="heading-53">How to use it</h3>I recommand using it in weight 0.6-0.7. If you have other loras used at the same time. Bring it to 0.7-0.8.This lora has two triggerwords. One is hyper_penis, use to generate penis without cum. The other is hyper_ejaculation. Use it when you need cum in your image.Why I made these two triggerwords? The answer is that, if your trainset contains both images with and without cum, the concepts will converge and you will get penis with cumdrip randomly. So I divided my trainset into two parts and made two triggerwords.<h3 id="heading-54">Training details:</h3>Traingset is of about 400 images. They are mirrored before training. Half of them are chosen as regularization images. Total steps are about 14000.Regularization = trueresolution=512batch_size=2epoch=10network_dim=32network_alpha=32clip_skip=2Using AdamW8bits as optimizer:<ul><li>lr="1e-4"</li><li>unet_lr="1e-4"</li><li>text_encoder_lr="1e-5"</li></ul>Locon parameters:<ul><li>conv_dim=4</li><li>conv_alpha=4</li></ul>

Hyper futanari (No Update)

In v2.0, I removed all the images with black bar censor and retrain this locon. That will do help to get better result. The annoying censor bar won't be there.I set batch size = 2 to acclerate training. The final result seems good.

Files with SHA256: b958b442c4eb8df2c04483dfbbf19fde394ed527580ced4eded0b0d75080067e