Anonymous
6/22/2025, 10:48:01 PM
No.105674559
>>105674431
You just set the dimension and alpha for the convolutional layers. I usually set them the same as the network dimension and alpha. You need to choose a non-standard lora type in order to use them.
Regularization images aren't necessary if the model is already familiar with the concept you want to train. In my experiments, what the regularization images did was acting like a combination of a controlnet and img2img, and not at all like what people say online.
If you're training for illustrious, you should use mostly danbooru tags.
You just set the dimension and alpha for the convolutional layers. I usually set them the same as the network dimension and alpha. You need to choose a non-standard lora type in order to use them.
Regularization images aren't necessary if the model is already familiar with the concept you want to train. In my experiments, what the regularization images did was acting like a combination of a controlnet and img2img, and not at all like what people say online.
If you're training for illustrious, you should use mostly danbooru tags.