search

cloneofsimo / minRF

Minimal implementation of scalable rectified flow transformers, based on SD3's approach
Apache License 2.0
317 stars 17 forks source link

minSNR loss weight? #2

Open Xynonners opened 2 months ago

Xynonners commented 2 months ago

Hi,

I'm wondering if minSNR is compatible with this approach.

zaptrem commented 1 month ago

They're more likely to sample timesteps from the middle which has a similar effect, I think. One of the SD3 paper authors told me they tried it at some point and it didn't make much difference.