lllyasviel / ControlNet

Let us control diffusion models!
Apache License 2.0
30.67k stars 2.75k forks source link

The discussion about the "sudden converge" #628

Open CuddleSabe opened 10 months ago

CuddleSabe commented 10 months ago

Hi, I found that u said there is a "sudden converge" I think that because of the uniform sampling of timestep in training. If use (large timestep, 1000) in the start, and it decay to the (0, 1000) while training the sudden converge is coming fast with small batch, such as 4.