*Possible important fixes for some issues with flux

1. is there something similar to logit normal? like the one in onetrainer?

2. Debiased noise estimation, Min SNR Support? do they work?

3. Noise offset? not sure if it works either.

4. Optional EMA, EMA decay?

5. Different learning rates for the t5 and clip L

6. Lokr training ( apparently it helps with concepts not getting bleed through)

Thank you for the work and support!

Hope at least some of these are possible in the near future.

Review the current Flux training support and compare it with the six requested areas: logit-normal sampling, debiased noise estimation, Min SNR, noise offset, EMA, separate T5 and CLIP learning rates, and LoKr training. Determine which items are already supported, which are feasible, and define a separate, testable scope for any accepted change.

Written by the indexing model from the issue text.