Trained on 768x768. Best merged with the 512x512 version.
These loras are mostly to be used to merge into other models using Kohya_ss script/bmaltais Kohya_ss gui. You can merge any lora you find to a model with the script/gui so experiment to get the type of look/composition you want in your final model.
The merging of loras work the same as models. The more you merge, the more 'muddy' the output becomes. You can think of it like mixing all the colors on a painter's pallet. It's better to have some restraint when mixing.
An example mix:
SD Model:
./Stable-diffusion/darksun_v60AkaFutasun.safetensors
LoRA model "A" - Model A merge ratio: 0.25
./stable-diffusion-webui/models/Lora/futanarimixV7-v1.1.2-Lora-DAdaptLion-D256-A1-000040.safetensors
LoRA model "B" - Model A merge ratio: 0.25
./stable-diffusion-webui/models/Lora/futanarimixV7-v1.1.3-Lora-DAdaptLion-D256-A1-000025.safetensors
LoRA model "C" - Model A merge ratio: 0.4
./stable-diffusion-webui/models/Lora/futanarimixV6-v1.1.9-Lora-DAdaptLion-D16-A8.safetensors
LoRA model "D" - Model A merge ratio: 0.25
./stable-diffusion-webui/models/Lora/futaveiny.safetensors
Out:
./stable-diffusion-webui/models/Stable-diffusion/futanarimix.safetensors
Description
FAQ
Comments (4)
holy H3LL
THX for the merge info. I am currently preparing a full dataset to train my own futa LoRA. I want to merge a checkpoint specialized in generating futa girls, but I hate that every LoRA out there requires "futanari" or even "1boy" as trigger word. So I'm gonna caption them with only "1girl" and "penis". Finger crossed that it works.
Anyways your LoRA seems like a good resource for me to get more image for my dataset. Im gonna try it out later. Cheers!
I'm looking forward to your new lora. I hope it goes well. If you figure out how to make a futanari lora without the style bleeding in from the training images, please let me know. I've tried everything I could to not have that happen and trained over 200 attempts.
One thing I learned was that you don't even need to use "penis" as a caption. That's why I chose something made up like "futanarimix" to describe the concept of futanari.
For tagging purposes, I remove any caption related to penis, testicles, futanari, erection. I append to the start of all captions my made up tag. Then I turn on shuffle captions and use keep N captions = 1. I believe this allows the lora to learn that the other parts of the image are not considered important to learn. That way the details of the penis will be better.
@screamlouder That's good to know! Thanks for you insight.👍