search

RohollahHS / BAD

The official Pytorch implementation of “BAD: Bidirectional Auto-regressive Diffusion for Text-to-Motion Generation”
30 stars 2 forks source link

About Diffusion #1

Closed feifeitu1 closed 2 months ago

feifeitu1 commented 2 months ago

Hello, I have read your paper and find your ideas to be very innovative. But I have a small question that I hope you can clarify for me. It is regarding the diffusion part — which section does it appear in? It seems that I didn’t see any description of the related content in the paper. Could you help me with this? Thank you very much!

RohollahHS commented 2 months ago

Hi, thanks for your interest. We refer to our permuted corruption process as a way of diffusion, similar to the paper Structured Denoising Diffusion Models in Discrete State-Spaces, which discusses masking as a diffusion process called absorbing diffusion. We can think of it as gradually corrupting the entire sequence in a dependent manner to a noise (in our case, multiple learnable absorbed states or mask tokens), in contrast to conventional diffusion processes, like Denoising Diffusion Probabilistic Models, where diffusion is applied to each dimension (each pixel) independently.

Our corruption process is discussed in Section 2.2 of our paper in the Corruption Process part.

feifeitu1 commented 2 months ago

Oh,Thank you very much for your reply.! I have a general understanding of your ideas now, and I will read that part of the paper in more detail.