TheFutonEra
This is my seventh and eighth major publicly posted LoRA. This is my first Z-Image Turbo and Klein 9b LoRA and using AI-Toolkit.
Please click on "Show More" under "About This Version" to get strength, sampler and scheduler recommendations. They are different for each model type.
This LoRA was an experiment to test the "wisdom" that a dataset should be "small and high quality", by training a LoRA on a large dataset of low quality images. The source images used in this LoRA were mostly from DVD frames from 2004-2011. Approximately 25% of images were from 2011 and high definition frames of that era. A total of 10956 images were used in the data set.
It is okay. It struggles in resolving a glans next to lips. Please continue reading for setting recommendations:
Summary for LoRA:
This is a concept LoRA that produces two men engaging in oral sex. The goals of this LoRA were to:
generate two men engaged in oral sex.
test whether using many low quality images instead of few high quality images was viable.
I consider this an alpha because like TwinkCockXL and with TwinCockFlux, it works most of the time, but still is not perfect..
The tags were generated by joy-caption-beta-one.
Tags were manually edited.
The primary activation tag "TheFutonEra" was added to all images. A secondary tag "1face" or "2face" was added depending on whether the image was a closeup only showing one face, or a slightly wider shot showing two faces.
Other notes:
Ethnic and age representation will likely rely on the main Z-image turbo model.
The length or size of the penis, and the circumcision status of the penis was not consistently tagged in this LoRA. Z-Image Turbo by default seems to favor distorted penises.
Women were not included in the training data, and I have no clear idea what will happen if a woman is specified, One likely result may be a penis being added to any figure in the generation. I do not plan on including women in the future as I do not have source images.
Regularization images were used. Extensive testing on style flexibility beyond photorealistic has not yet been conducted.
Special thanks to @markury and the members of the Bulge Discord server https://thebulge.xyz for their support, advice, and beta testing.
Description
Initial Z-Image Turbo LoRA trained to 48k steps. Block 1 contains 1face and 2face images. In Civitiai: Choose either 1face or 2face, but not both.
Sampler, Scheduling, Clip notes for ZiT:
Initial testing was done on the Stability Matrix implementation of Forge UI neo. Testing was done using the z_image_turbo_bf16 model.
Recommended Shift is between 18 and 24.
Recommended Sampling steps is 18
Recommended LoRA strength is between 0.44 and 0.66.
The LoRA was tested mainly on Euler a and Linear Quadratic, Beta, or DDIM. Euler and Laplace also was tested.
SeedVarianceEnhancer Integrated was set to 2 steps with a Cosine decay.