The scale of timesteps in the text modeling examples

[realfolkcode]

Hi! Really enjoying the code and the paper. These are the most comprehensive resources on flow matching.

I noticed that the text modeling example might contain a subpar encoding of timesteps. In short, it uses the default max_period parameter from the GLIDE repository, which is a diffusion model with discrete timesteps in the range {0, 1, 2, ...}. However, as I understand correctly, in your example, all timesteps are sampled from the [0, 1] time horizon. See the similar issue that I've opened in the SiT repository. I think the most elegant solution is to rescale the timesteps similar to how Flux handles them.

Let me know if I am misunderstanding something, as I am just on my way of learning the flow matching framework.

[itaigat]

Thank you! we were not aware of this, we are experimenting now with time range [0,1] and period of 0.001 and 0.005. We will update.