search

paTRICK-swk / D3DP

[ICCV2023] The PyTorch implementation for "Diffusion-Based 3D Human Pose Estimation with Multi-Hypothesis Aggregation"
MIT License
158 stars 8 forks source link

About the reverse process #6

Closed WilliammmZ closed 1 year ago

WilliammmZ commented 1 year ago

Thanks for your interesting work~

I would like to know some details about the reverse process. Why did you consider applying the one-step solution for the reverse process? In my opinion, the multi-step method can produce higher quality results for generation tasks.

raulTrial commented 1 year ago

Hi,

Did you figure out why it uses one-step solution?

harryzhangOG commented 7 months ago

did you figure out a way to do multi-step denoising? thanks.

WilliammmZ commented 7 months ago

did you figure out a way to do multi-step denoising? thanks.

In D3DP, the train process is DDPM and the inference method is DDIM. You can get the way to do multi-step denoising or one-step denoising in the paper called DDPM and DDIM.