An example showing hands in basically the easiest, “flattest” pose you can get, still failing.

Do diffusion models find this easier than other poses?

Still several times better than before...

Looking forward to mj six! Giving the doubters the finger edition

I see you have six fingers on your right hand. Someone was looking for you.

A person with six fingers? That's inconceivable!

How does one give the middle finger with six fingers?

Twice as well as someone with five? I guess you could do some sort of hybrid British/American gesture with two middle fingers.

It’s perfect for Aussie-Americans- they can give the forks and TWO middle fingers simultaneously!

It's a bit better than before but still not right:

https://twitter.com/Excaldata/status/1636375182750396418?s=2...

Now, I'm traumatized.

probably a 5-10% error rate. C'mon, you have to be impressed though compared to where we were only a few months ago

I'm extremely impressed compared to where it was before (which was frankly frightening) but the error rate is pretty bad really, if a person made this you'd either assume it was intentional or they had severe problems. It may be 5%-10% looking at individual fingers to see if they are correct but each hand has (somewhere around) 5 fingers to get right. Diffusion just doesn't lend well to connected things like counts of objects or coherent text.

It's AI, you, have to be impressed right?

Well I for one, as a person with sophisticated tastes, find the emperor's clothes to be absolutely fabulous, none finer have I ever gazed upon.