Finetune details
#5
by
bielca98
- opened
Hello,
I was wondering which parts of the model were exactly fine-tuned. Did you only fine-tune the VAE and keep the backbone model unchanged, or did you also retrain the UNET?
Thanks.
It's actually the other way around. It uses the same VAE used in stable diffusion 2.1 and the unet was tuned for 300,000 steps.
There are some basic details in the model card:
Training details
All parameters were initialized from the stabilityai/stable-diffusion-2-1-base model. The unet was fine tuned as follows:
U-net fine-tuning:
- 200,000 steps, learning rate = 1e-5, batch size = 992 (248 per TPU).
- 100,000 steps, SNR gamma = 5.0, learning rate = 1e-5, batch size = 992 (248 per TPU).
- Trained on LAION Improved Aesthetics 6plus.