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
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.
Comments
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.
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
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.