teapearce / conditional_diffusion_mnist Goto Github PK

Hi,
If we parameterize xθ in terms of x0-prediction, how to implement the inference of classifier-free guidance? The original paper only parameterizes xθ in terms of ε-prediction.

Hello author, your work is very good, but the sample speed is a bit slow. I would like to ask if it is possible to provide ddim code?

Where is the “Puncond”in the original paper, please!

Hi,

Thank you so much for the code!

I am working with this code these days and I found that sampling from a trained model consumes significant memory. I looked into this problem and I found that it is helpful to add a line "x_i = x_i.detach()" at the end of the for loop in function DDPM.sample in script.py.

I hope it helps!

Thanks,
Harry

Hi,

Thanks for your code. I have one question about the reason why we need to caculate 'z' caculated in line 284: z = torch.randn(n_sample, *size).to(device) if i > 1 else 0.

Hello, thank you for your excellent work!
I am a green hand in diffusion model.
If I want to further improve the performance of model, what parameters and network structure do I need to modify? Do you have any suggestions? Thank you!

do you have the weights of a trained model to share?

Hi
Thanks for this code.
What you suggest for evaluating this conditional generation?
E.g. what is the easiest code change to get a meaningful evaluation with some metrics (e.g. what about FID)
Any suggestion on which metric could be suitable here

Best regards

why it is context_mask = (-1*(1-context_mask)) # need to flip 0 <-> 1 not context_mask = (1-context_mask)) ?