Z-Image Base v3
Trained with over 700 images for 30,000 steps, and this is my best looking one yet. A new model will probably dethrone this as I am writing this model though. It was looking good at 20,000 steps but since the z image turbo LoRA took 30k steps to look good I ran this one for the same 30,000 steps, but at a lower learning rate for the last 10k steps to prevent overcooking.
Z-Image Base v2
Trained with over 700 images for 20,000 steps
This version should help some with the glans direction issue but it keeps creeping up about 25% of the time wrong head direction. Prompting around that can help either with positive or negative prompts if you have cfg > 1 which you should be running this model at.
Please share your work or experiences, otherwise not much incentive for me to keep sharing these. Quid pro quo Clarise.
Check out coachbate.ai if you like my work and want to see more.
Z-Image Base v1
trained with Z-Image base model using 150 images and 25 regularization images using AI Toolkit for 10,000 steps.
Need to reduce strength < 1 if using multiple LoRAs just like the Turbo version.
Finding great results with LLM writing prompt and sampler res_2s, seed_v2 and seed_v3, and res_multistep. Clownsampler with the bongmath also gave better image than standard KSampler which produced monsters I didn't get as often with the ClownSampler
Description
Z-Image Base v2
Trained with over 700 images for 20,000 steps, and this is my best looking one yet. A new model will probably dethrone this as I am writing this.
This version should help some with the glans direction issue but it keeps creeping up about 25% of the time wrong head direction. Prompting around that can help either with positive or negative prompts if you have cfg > 1 which you should be running this model at.
Please share your work or experiences, otherwise not much incentive for me to keep sharing these. Quid pro quo Clarise.
Check out coachbate.ai if you like my work and want to see more.
FAQ
Comments (3)
Training a penis lora is so difficult! I am training my own right now and creating 6 different versions to test. Did you caption your images? I am using a dataset of 40 captioned images. All images created with Chroma because I love that models penis the best. Great job on this btw!
Yes captioned and captioned. Still wish I had the exact formula but it’s a little luck and model dependent. Klein 9b seems like it tries to understand anatomy the best - look at my video for that model for the steps progression of the evolving penis
For character Lora‘s with the latest models you’re supposed to caption everything that you don’t want in your Lora so include the background , the outfits but don’t say muscular or skinny. Don’t say the eye color .for a penis Lora you can’t really do that cause the model doesn’t know about penis so I have lots of detail in there a couple times i had really specific anatomy detail in there and that it seems to help. Sometimes I just say cut uncut hard soft big small above average.