So I published my 100th LoRa yesterday, but since I was too tired I did not write an article. So this is the accompanying article, I’ll keep it short. If you are curious what it is, here you go.
When I started publishing LoRas civitai was merely like a repository for me. I was satisfied with people downloading and using my models.
However that has changed during the last couple of months and I really felt like becoming a part of the community. So I want to say thank you to the people I talked to, people I interacted with, people urging me on and people using and taking care about my models.
Actually I did not plan do something special for my LoRa number 100 at first. But because Nanashi made a massive effort with the 200th LoRa and Citron told me number 100 is something special, I thought I’d also have to do something special. It turned out to be a lot more work than I initially thought it would be (nothing unusual here).
I rummaged through four seasons of screencaps and additional images and had 1.381 raw images in the end. After selection I still had 652 images that went into the training data. I cropped every single image by hand. I used an autotagger, since I have gotten used to it, but that does not mean the work is done. I’d still have to work through every subset of the data and apply the correct trigger words, prune unnecessary tags and apply correct outfit tags.
I made several test training runs of data subsets to see, if they would work as expected.
To avoid overfitting and for better details I tried a new training approach. Instead of a high learning rate, I used a moderate learning rate with more training steps in the first run.
Additionally I used the training weights of the first run as basis for a second training run with low learning rate to train in details and improve concepts that were undertrained after the first run.
In the end I had a total of 44.740 training steps and more than 30 hours of training time on my gpu. I guess I could have done a third or fourth training run, but I am not sure, if I’d get better results in the end and maybe it’s already crazy to do something like this anyway.
I put a lot of work into this LoRa and I would appreciate it, if you give it a try or check it out.