"PNTE" Negative Stable Diffusion LoRA
Increase the quality and amount of details in images with these negative LoRAs for Stable Diffusion models.
How to use:
THESE ARE MEANT TO BE USED WITH NEGATIVE STRENGTH VALUES.
To do this simply add the LoRA to the negative prompt or manually adjust the LoRA strength depending on the diffusion interface you are using.
For ease of use, ComfyUI users can use the SnakeOil custom_node suite I created. This node not only automatically inverts negative LoRAs but it also looks for them in models/nloras folder rather than the normal models/loras folder. This helps with organization and keeping our ever growing lists of LoRAs a little shorter.
SD3.5L:
The most recent updated version my "point-e" negative embedding for use with Stable Diffusion 3.5 Large was trained with my experimental custom config for aitoolkit.
I set CLIP to 0 as I did not train the text encoders.
LoRA strength can range from -0.01 to -2.00 but there will often be deterioration past -0.75 in most cases. -0.45 is a good place to start. Works as low as -0.01.
Description
Trained with my experimental aitoolkit config file. See article below for more information.
This works very well as an incremental detail slider. I normally start around -0.45 for overall better results. Things can get a little strange the closer to -1 you get and there are at times a noticeable shift between -0.78 to -0.98. Use as low as -0.01 for extremely subtle quality increases.
This version was trained with a network rank of 64