A LoRA Slider collection, including age_slider, weight_slider and so on.
LoRA weight can be set from -10 to 10 depending on your case.
Recommand using sd-webui-ctl (an extension of A1111 webui, https://github.com/cheald/sd-webui-loractl.git) to control lora weight as 0 at first few steps in diffusion progress, for achieving fewer changes in background.
Training script has been released (https://github.com/HaohuaLv/LoRA-Slider.git).
Lora-Slider for SD1.5 and SDXL have been released.
Description
LoRA weight: -10 ~ 10.
Negative LoRA weight: female.
Positive LoRA weight: male.
FAQ
Comments (8)
I can see the other two LoRAs are rather subtle, this one tho... oddly I found the male-female cutout value at 2 for some reason. Test on the ICBINP model btw
P.S.: also, the gender slider doesn't seem to work well with east-asians (only tried with myself so I dunno may be an outlier)
For the male-female cutout value at 2, I think it may be due to used prompt's potential tendency.
For the gender slider doesn't work well with east-asians, maybe due to that this slider model just trained on general trigger word like "person" "human", and no targeted training for east-asians.
A new gender_slider-asian model has been released. Maybe it would achieve a better performance for east-asians.
Looks great, but in your examples the subject changes from Asian to Latina.
A new gender_slider-asian model has been released. Maybe it would achieve a better performance for east-asians.
Finally a slider to turn Elon Musk from a weird looking man into a weird looking woman!
Elena Musk
Thank you for your very meaningful project. I have some questions to ask. How many pieces of data are needed to train an age slider model What size is generally suitable for setting the Lora Rank.