I trained this model based off the awesome instructions from Aitrepreneur's Youtube Video. I had to change a few settings from the Youtube instructions from Aitrepreneur's video. Namely: Mixed precision and saved precision I used fp16 instead of bf16, and I unchecked Use 8bit adam in the Advanced Configuration section. Otherwise I'd get the CUDA out of memory error while training. Also, if you use his configuration file with Kohya, you'll have to first update your file paths under the folders tab. I used 30 images of my dog and found that a higher number of images produced better results in my case. I also prefixed all of the blip captions with the name of the model (Malu).
In the example outputs shown here, I used the Rodent Diffusion 1.5 model. Very much open to questions and or suggestions / constructive criticism. Long live stable diffusion and the stable diffusion community!
I trained this model based off the awesome instructions from Aitrepreneur's Youtube Video. I had to change a few settings from the Youtube instructions from Aitrepreneur's video. Namely: Mixed precision and saved precision I used fp16 instead of bf16, and I unchecked Use 8bit adam in the Advanced Configuration section. Otherwise I'd get the CUDA out of memory error while training. Also, if you use his configuration file with Kohya, you'll have to first update your file paths under the folders tab. I used 30 images of my dog and found that a higher number of images produced better results in my case. I also prefixed all of the blip captions with the name of the model (Malu).
In the example outputs shown here, I used the Rodent Diffusion 1.5 model. Very much open to questions and or suggestions / constructive criticism. Long live stable diffusion and the stable diffusion community!