DALL-E 2 is stunningly better than DALL-E 1, but what's interesting (and underappreciated) is the fact that the newer model is much smaller & amount of training compute is similar. Improvements are essentially all due to algorithmic innovation.

May 23, 2022 路 8:42 PM UTC

22
64
6
943
Load newest
Replying to @gdb
Impressive 馃憤
Replying to @gdb
Is the deep learning sign bug fixed as well? And on a side, is this possible "movie like poster picturing in a dramatic way two hitmen shooting at each other "? Please.
Replying to @gdb
Don鈥檛 you also use a different dataset? This might be a big one
1
Replying to @gdb
Can confirm it's super fast. Keep up the great work.
3
Replying to @gdb
What if you applies 2鈥檚 algorithm to 1鈥檚 parameters
Replying to @gdb
Does that include algorithmic innovation in how the training data was selected (even if by volume it's the same size as before) or do you mean only "inside the model"?
1
Replying to @gdb
decoder as diffusion model is it ?
Replying to @gdb
Hey Greg by chance may I request an AI piece to be done?
Replying to @gdb
JUKEBOX 2 then...
3