differences on sampling between the paper and your implementation about denoising-diffusion-pytorch HOT 4 CLOSED

Lucidrains' implementation is (predicted noise -> x_0 -> x_t-1), which uses eq. 9. Routine 1 (predicted noise -> x_t-1) is eq.10. I've verified and they are mathematically equivalent (hint: plugging eq.7 into eq.9).

from denoising-diffusion-pytorch.

@askerlee Thanks for your help

from denoising-diffusion-pytorch.

Lucidrains' implementation is (predicted noise -> x_0 -> x_t-1), which uses eq. 9. Routine 1 (predicted noise -> x_t-1) is eq.10. I've verified and they are mathematically equivalent (hint: plugging eq.7 into eq.9).

I switch the sampling method from eq.9 to eq.10, yield very poor quality and unstable samples.
I completely follow "Algorithm 2 Sampling" from the paper.

Update:
Is the clip op, remove x_start.clamp_(-1., 1.) in the implementation, It will also produce bad quality sample which "Algorithm 2 Sampling" produced.

from denoising-diffusion-pytorch.

@bot66

Update: Is the clip op, remove x_start.clamp_(-1., 1.) in the implementation, It will also produce bad quality sample which "Algorithm 2 Sampling" produced.

Do you have an idea of when / in which conditions x_start.clamp_(-1., 1.) should be used/not be used?

from denoising-diffusion-pytorch.