Why use x_start as the target in each timestep of diffusion training?

Stanford-TML / EDGE

Official PyTorch Implementation of EDGE (CVPR 2023)

https://edge-dance.github.io

MIT License

454 stars 68 forks source link

Why use x_start as the target in each timestep of diffusion training? #35

Open shaoguowen opened 10 months ago

shaoguowen commented 10 months ago

I have seen using noise, x_noisy or v_prediction, etc. as the training target, but each timestep uses x_start as the training target, which seems a bit strange. Can you explain it or provide relevant articles?

plyfager commented 8 months ago

Maybe it's convenient to calculate other losses.