I would be interested in reading a scientific paper ablating reasons for how AI got hands/fingers right.

@_arohan_ 1. More scale 2. More data 3. Better understanding of data mixes

my intuition for why hands required more scale/data/targeted collection was a combination of (1) hands are often occluded, fingers wrapped or hiding behind each other and (2) they don’t have horizontal symmetry, which partially explains why other things like people with three arms were rarer

@_arohan_ During Uni-1, we unironically had a very very tough hand prompt that i would personally track over time of an image that had 3 hands holding each other (two adult and a baby hand). It was genuinely garbage for 99.99% of training steps and then locked tf in at the end

@mo_norouzi is this a joke? each hand has like 12 fingers

@_sam_sinha_ @_arohan_ Increasing resolution helps a lot too. Detailed descriptions make the task easier. We don’t have a systematic study for this unfortunately. It can make a good research paper with title “What Solved the Finger Count Problem in AI Image Generation”

@_sam_sinha_ Is there a log compute plot for this? When do we get to 5 fingers

@jxmnop

@mo_norouzi 👌

@mo_norouzi 🤔

@mo_norouzi nearly 🫰 made in @ideogram_ai

@mo_norouzi Now it’s just perfect 🤌💎

@mo_norouzi I loved @ideogram_ai from the first minute

@AmirMushich very realistic skin texture.

@NimEshed @ideogram_ai how can we make it better?

@_arohan_ pretty sure it's just dataset. are you really sure you are getting diverse hand/finger positions? they just overtrain on hand datasets lol

@mo_norouzi We're right there with you! created w @ideogram_ai

@mo_norouzi It is very important that you *not* let David Cronenberg discover your model. 🤪

@mo_norouzi @_arohan_ Increasing resolution ofc helps a lot but tracking it along a constant resolution was pretty interesting.

@mo_norouzi eww i love it!