RE: AI generated images from my custom diffusion models
You are viewing a single comment's thread:
Hi @lavista, Thx for your experimental notes about the training. I may have missed the info, but how many pictures did you gather to make model ? What were your selection criteria for a picture to make part of your dataset (beside HQ size) ? Is colab pro enough ? what was the processing time ?
@Kaliyuga PAD and now your model are opening new mind bending levels a latent space exploration. This is magic.
0
0
0.000
Hi @dbddv01! I duplicated most of the images by flipping them horizontally, so I had a total of 132 images.
Regarding the selection criteria, it depends on what you want to do from the model you create. I wanted the children’s illustration / fantasy kind of look. So I choose images accordingly - basically most of my wife's work. Also, I tried to select mostly square format images. I do have some rectangular images in my dataset, but I believe the training program crops the non square images.
Colab Pro is enough, I did it with Colab Pro. My first model I trained for several days, the training steps count was 152000 when I stopped. With my second model, I started getting decent results at 22000 steps and I stopped training at 32000 steps as I liked the results already. This took only a couple of days of training- I was not training 24hrs/day. Say, I used to train for around 7-8 hours per day for 2.5 days, that's around 24 hours of training for the second model to get good results.
To be honest, with my first model, at around 10000 steps, I wanted to see if it will produce something. I ran it and this was the image produced:
It indeed felt like magic and I was so satisfied and happy!:).
Looking forward to what you come up with now!
Thanks a lot for those details. It's looks technically quite affordable, I will have to think of it now. This is exciting !
I'm just butting in to say that more images are probably better! I'd recommend 1000 or more so that you don't overtrain :)
Definitely. More the better:)