HOME | DD

Description
~Nyanna Mew LoRA~

Trained an SD v1.5 LoRA for my 3D "Nyanna Mew" catgirl using only 47 images. In theory, I should had used 100-300 references, but I wanted to know if I could get away with less than 50. I can, but the training overfits with more than 2 epochs. As a result, the model works, but it's not as versatile. I have to keep the model weight under 50% to use it. That's what happens when we lack enough images for proper model training!

However, I was still able to obtain these cute Nyanna Mew renders with the usual prompt engineering. The training has captured her looks and detailed outfit quite well! So YES, we can get away with less than 50 images, but we pay with lack of LoRA versatility. I have tried several different ways to train the model, but there was no escape from the need for more references. Simple flat-shaded cartoonish or Anime characters are fine with fewer images, but not something like this. Lesson Learned!

LoRA trained over 47 images rendered in Poser using 50 steps over 3 epochs in batches of 4. In spite of the overfitting, the training ended with less than 10% average loss, which is pretty good! Character made in Poser, LoRA trained in Kohya_SS, rendered in Stable Diffusion, postwork in PaintShop.

Description
~Nyanna Mew LoRA~

Trained an SD v1.5 LoRA for my 3D "Nyanna Mew" catgirl using only 47 images. In theory, I should had used 100-300 references, but I wanted to know if I could get away with less than 50. I can, but the training overfits with more than 2 epochs. As a result, the model works, but it's not as versatile. I have to keep the model weight under 50% to use it. That's what happens when we lack enough images for proper model training!

However, I was still able to obtain these cute Nyanna Mew renders with the usual prompt engineering. The training has captured her looks and detailed outfit quite well! So YES, we can get away with less than 50 images, but we pay with lack of LoRA versatility. I have tried several different ways to train the model, but there was no escape from the need for more references. Simple flat-shaded cartoonish or Anime characters are fine with fewer images, but not something like this. Lesson Learned!

LoRA trained over 47 images rendered in Poser using 50 steps over 3 epochs in batches of 4. In spite of the overfitting, the training ended with less than 10% average loss, which is pretty good! Character made in Poser, LoRA trained in Kohya_SS, rendered in Stable Diffusion, postwork in PaintShop.