Question about generation process
Hello!
These are fantastic. I have only just started testing building custom models and this was the first LoRA one I have seen. I have seen you comment elsewhere you don't share your training images. But I thought I might ask how many images you used and where you sourced them? How many attempts/tries did it take you to arrive at these results?
Image count doesn't really matter, unless it's too low. I can get good model with 40 images or with 3000. Done like 12+ tests on various datasets/concepts to find every potentially good setting and I just train with the one that works in almost all cases. If the results aren't great for my liking, I just change settings and try again. If they don't work, my dataset sucks. You can get the images from any source you want really. Most LoRAs here are first attempts, I had a bit of practice with model finetuning/dreambooths so I already know what to change/look for or how to train. With this amount of LoRA. I'm not chasing perfection like most people that will do 1 or 2 of them. I just go with 80/20 rule for most of them.