Lesson 21: Deep Learning Foundations to Stable Diffusion

[@mathumble6684]

1:07:12 I think it is because we optimize against a score, and the arg_max/ arg_min does not change when the statistical ranges are offset or scaled, e.g. optimizing x-1 is the same as optimising x skipping sampling process basically is similar to the linearization (EKF) / momentum, linearly extrapolate using the result of previous time step

[@coolarun3150]

so far detailed and awesome!!!

[@howardjeremyp]

Glad you think so!

[@michaelmuller136]

Very interesting, thank you!!

[@bayesianmonk]

Sometimes explaining the math helps more than escaping it, no heavy math is used anyway. I found the explanation of DDIM not very clear. Thanks for the course and videos.

[@thehigheststateofsalad]

We need another session to explain this process.

[@maxkirby8500]

Yeah. I've been spending quite a bit of time trying to bridge the gap by reading through the papers and stuff, but maybe that's intented...

[@satirthapaulshyam7769]

Samples in these diffusion models r b2in -1 and 1 29:59

[@kettensaegenschlucker]

1:27:35 - Wondering what made you cheer, Johno ... 😂 Edit: spoiler alert, but there is a resolution shortly after

[@frankchieng]

i thought in the class of WandBCB(MetricsCB) def _log(self, d): if self.train: should be modified with if d['train']=='train'