Does anyone have a tutorial on how to utilize LORAs?
I'm trying to learn how to use LORAs with existing models before I attempt to make my own LORAs and models. I downloaded a few LORAs (MonkeyIsland_v10 and VioletEvergarden_v10), and I have the protogen photorealistic and anime models.
I then installed the Kohya-ss additional networks extension from the StableDiffusion WebUI and restarted the program.
At this point I can see both of my LORAs in the addional networks section and I can click on them to get the LORAs added to my prompt.
My issue is that when I attempt to generate an image that looks good without the LORA, for example:
"Extremely detailed, photo of a pirate, flintlock pistol, island"
it generates an image that is distorded (kinda looks like a stained glass image) and does not apply the intended effect very well.
I repeated this with both protogen models and with both LORAs.
My lora prompts looked like this:
"Extremely detailed, photo of a pirate, flintlock pistol, island, <lora:violetEvergarden_v10:0.8>"
no matter what value I placed for the LORA I could not get images that looked like the intended effect. I could see the effect trying to be applied, I could see some resembles. Is this because the LORA strenght/weight is too low? can I go to values above 1?
any ideas and suggestions or tutorials would be greatly appreciated!
oh I forgot to mention I was using the EULER A sampling method, sampling steps of 50, and CFG scale of 12.
If you're using the Kohya-ss additional networks extension then you shouldn't add the "<lora:lora_name:1.0>" to the prompt. Just select the LORA and adjust the sliders, and make sure to check the "enable" checkbox.
Try going back to the LORA web pages and copying a prompt/settings and try duplicating a sample image. Once you have it working as intended by the uploader, you can tweak from there. Sometimes LORAs aren't super flexible and only work best with certain models, so also pay attention to what the uploader recommends there.
You can use LORA weights up to 2, though most models will "burn out" at that strength.
if its a normal lora just use this in prompt <lora:model_name:1.0> and try reducing the cfg scale .and dont use both in prompt and Kohya-ss additional networks together. use one on the other
Also don't fry the config scale on your Lora to past 1, dont' be me don't forgoet to lessen the weight and instead have a CONFIG WEIGHT OF LIKE STRAIGHT 8 - yo uget fried eggs for breakfast.