the distribution of t

[felixdae]

x-flux/train_flux_deepspeed_controlnet.py

Line 219 in 4749542

t = torch.sigmoid(torch.randn((bs,), device=accelerator.device))

as this line show, why we use a Logit-normal distribution for t here.
from the flow matching paper https://arxiv.org/abs/2210.02747, t is sampled from uniform distribution U(0,1)