I read paper “lucidating the Design Space of Diffusion-Based Generative Models” and find that in original inference section, the σ is sampled according to the ρ, σ max and σ min but in diffusers the code in scheduling_euler_discrete.py shows that the σ is sampleing according to timesteps, betas and alpha.The difference above also influences the training section. could someone give me some explanation？ Thanks a lot
I got it, the inference section is the same with the origin EDM