HN Explorer

6 comments

[ 2.2 ms ] story [ 26.4 ms ] thread
GaggiX
It's pretty cool to play around with the precomputed generations, it let you develop a better intuition around the samplers.

Like how non-ancestral samplers converge to the same image.

hmcq6
Agreed. Animations definitely helped. I noticed that more "errors" with wide landscape images than tall portrait ones. I have to imagine that's because the dataset is pulled from the web and the web has a preference for vertically segmented design
GaggiX
The SDXL model was trained on different aspect ratios, but each one of those aspect ratios has a precise resolution, so anything other than those resolutions would be considered out of distribution, so maybe when you were inspecting the wide landscape images you use an out-of-distribution resolution while with the portrait ones you didn't as much.
whywhywhywhy
This is a very limited "Settings Guide" there is a bunch of stuff they're not even mentioning, maybe not aware of. Such as the fact it now has 2 positive prompts + 2 negative prompts and tweaking the amount of the refinement.
anotherjesse
Agreed! We wanted to share what we had today and iterate.

We have a section at the bottom on "Ensemble of experts" - allowing you to change the percentage run with the base model vs the refiner.

I am most excited to add the ability to change between prompts