Hello, thank you for your open-source work, I have gained a lot from it.
When I try to train VQGAN from scratch to reconstruct plant300K data, I find that the reconstructed images are quite blurry, shown as below:
I didn't change the default hyperparameters (and the default model is VQGANLight, I didn't change it as well), and the loss curves are a bit weird, shown as below:
With this VQGAN pretrained model, I failed to train a good DDPM model.
I wonder that if you can reach a better VQGAN result, and what is the corresponding hyperparameters?
Hello, thank you for your open-source work, I have gained a lot from it. When I try to train VQGAN from scratch to reconstruct plant300K data, I find that the reconstructed images are quite blurry, shown as below: I didn't change the default hyperparameters (and the default model is VQGANLight, I didn't change it as well), and the loss curves are a bit weird, shown as below:
With this VQGAN pretrained model, I failed to train a good DDPM model. I wonder that if you can reach a better VQGAN result, and what is the corresponding hyperparameters?