Open JunMa11 opened 2 years ago
Also waiting for their response (if there will be any),
The weights for SR diffusion is only 455mb (after clean up and separate diffusion part & image enc/dec part; smallest among txt2img(3gb) and infilling(1.55gb)), should not be such a vram monster to train imo.
And if it can directly use SD's first_stage_model
weights (means you don't need train a image enc/dec yourself), and works under [1,4,H>>3,W>>3]
latent,
it would be even lighter to infer & train.
@JunMa11 The notebook is strange and does not seem to correspond to the results from the paper. Do you know why is it doing 256->1024 and not 64->256? Do you know what data was it trained on? If you could figure out how to make it work for 64->256 I'd be happy to hear.
Thanks!
Dear all,
Thanks for sharing the great notebook for SR model inference: https://colab.research.google.com/drive/1xqzUi2iXQXDqXBHQGP9Mqt2YrYW6cx-J?usp=sharing
Could it be possible to share a notebook for training the SR model?
I am also eager to have a look at that notebook!
Dear all,
Thanks for sharing the great notebook for SR model inference: https://colab.research.google.com/drive/1xqzUi2iXQXDqXBHQGP9Mqt2YrYW6cx-J?usp=sharing
Could it be possible to share a notebook for training the SR model?