Dazu kommt wohl die Tatsache, dass die Schweiz nicht Teil der NATO ist (als einziger A-Tier-Staat ist Taiwan nicht mit den US verbündet - aber in Taiwan werden Chipkomponenten hergestellt)

Also, the code for creating the models and datasets can be found here (in case you need a hint): github.com/mattminder/l...

I would be very interested to see how their learning curve compares to „regular“ SGD

Ah I didn‘t consider the number of diffusion steps. It would be interesting to investigate the interaction of the number of steps and guidance… Anyways, thanks for the response and congrats on the nice results!

What I’m confused about: In image diffusion, lower guidance is associated with lower FID scores and thus better similarity between output and true distributions. So to me these „outliers“ are unexpected even at low guidance. But with higher guidance your model performance is impressive :)

On some toy examples that I‘ve tested, I get quite a bit of points being predicted either in the ocean or in other unlikely places like Antarctica. To me this looks like the model hasnt yet learned a good posterior distribution on the globe. Do you monitor these unlikely predictions during training?