drhead

> I’ve been seeing much closer prompt adherence especially with regard to fine details after applying this patch. > > I haven’t used the new noise scheduler yet so is...

I think I have gotten it to work much more closely to what the paper authors demonstrated:   The desaturation is actually far less of an issue if you...

I've been experimenting with a model trained on v-prediction with zero terminal SNR, including with K-diffusion samplers. I've also applied the modification to the DDIM sampler under LDM and tested...

> > I think this does need implementation in the training examples and further testing before merging, including profiling performance gains vs. standard for loop implementations and versus regular blocking...

So early testing on this version is showing that the `foreach` implementation is visibly faster even without having run a profiler on it, and doesn't seem to use any more...

> Yeah would love to know the differences. I think not using `accelerate` is the primary difference here but of course, we need to know more and better. This appears...

After testing/profiling on a different machine I'm fairly confident that the non-blocking offload is working as well as a given environment/other parts of the code can permit it to work....

> > and the .item() calls on metrics being reported > > How significant is this one? They're not a huge issue if it is being called at most once...

> I assume not. But I will defer to @drhead for confirming. It's effectively complete except for linting

> This is good to merge. We'd need to: > > * Fix the code linting issues. > > * Add a block about this feature in the README of...