duanyiqun/DiffusionDepth

Question about DDIM loss

Opened this issue · 2 comments

zyp-byte commented

Hi, thanks for your wonderful work!
When I was reading the code, I noticed that you took the time embedding on the feature extracted from the RGB images. I am wondering if it is better to take the time embedding on the depth output by the decoder (namely 'refined_depth' defined in your code), or just annotated depth with masks.
Thanks for your work and codes again!

duanyiqun commented

Hi, thanks for your question.
At that time our thinking is more like to add the time embedding on a dense and consistent feature. This way is closer to the original diffusion model. I haven't try to put time embedding directly on depth map.
Have you got any attempts on that? If the results are positive, I'm keen to have a improved version with you.

VLadImirluren commented

Hi, thanks for your question. At that time our thinking is more like to add the time embedding on a dense and consistent feature. This way is closer to the original diffusion model. I haven't try to put time embedding directly on depth map. Have you got any attempts on that? If the results are positive, I'm keen to have a improved version with you.

Hi ! Thanks to your nice job.
I notive that you choose to predict the x0 instead of noise like DDPM.
Can you share the reason with me?
Thanks again~