It's sort of surprising how far behind real 3d geometry generation is compared with raster generation from diffusion models. As this doesn't feel like a particularly hard problem compared with say MJ 5, I sort of suspect the lack of performance in 3d generation is because most AI researchers do not understand 3d geometry well enough to have smart ideas for an approach to ml model design, and I would guess specifically struggle for smart ideas for a loss function framework (probably the most critical), so they've probably only tried a million or so model ideas, rather than the billions of attempts at 2d that have led to modern diffusion techniques through much more trial/error.
How long did you have to wait for your invite? I've been on the imagine waiting list for nearly a month now
The waiting list can take a long time, but I am not sure what you should expect. Sorry!
I'm curious what it could do with a (Image reference to 3D Model)
Currently it only supports text prompts. However another startup kaedim3D does this!
@@wintorartour I think that kaedim3D is a scam, th-cam.com/video/wB_NudGKKxI/w-d-xo.html th-cam.com/video/dnwDPLLzzEU/w-d-xo.html
A scam! 😅😅 oh noooo haha.
How long did it take to generate the model? A few minutes?
Too long ;) No just kidding, but about 30 - 40 minutes :o
It's sort of surprising how far behind real 3d geometry generation is compared with raster generation from diffusion models. As this doesn't feel like a particularly hard problem compared with say MJ 5, I sort of suspect the lack of performance in 3d generation is because most AI researchers do not understand 3d geometry well enough to have smart ideas for an approach to ml model design, and I would guess specifically struggle for smart ideas for a loss function framework (probably the most critical), so they've probably only tried a million or so model ideas, rather than the billions of attempts at 2d that have led to modern diffusion techniques through much more trial/error.
I am also surprised it doesn’t to better. But of course we should not compare 2D generation with 3D generation: it’s much more complicated
true
seems kinda fishy to meeee
🐟 Does it?