imagen-pytorch
imagen-pytorch copied to clipboard
Reducing ElucidatedImagen sampling steps
Hello, I am testing unconditional ElucidatedImagen
with num_sample_steps = (64, 32)
. I am wondering if it possible to reduce num_sample_steps
for base unet and super resolution unet to 8 or 4 while maintaining quality/diversity?
Is this, https://nvlabs.github.io/denoising-diffusion-gan/, applicable here to reduce the steps to 4 for faster processing?