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 · 8:42 PM UTC

22
64
6
943
Replying to @gdb
I wonder if you guys are planning to release—or make available somehow—the CLIP model. I know there's a "baby" version of the model (which I'm using) and it's darn good, so I can only imagine how much better the full version is.
6
Replying to @gdb
Google just released imagen. Which is better in your opinion @then_there_was
1
Replying to @gdb
hell yeah. and with all those savings you could make an even bigger model
1
15
Replying to @gdb @Plinz
what are the algorithmic improvements?
Replying to @gdb
Fascinating. Can you share some of the more significant changes?
Replying to @gdb
What’s really interesting is that DALL-E 2 came only one year after v1. Considering it’s already as good as some of the best human artists (and even better in some respects), it’s hard to imagine what v3 or even v4 look like.
2
12
Replying to @gdb
Exciting stuff!! Just filed for the Dall-E access. Thrilled to try it out! 🙌🙏
Replying to @gdb
Progress
Replying to @gdb
Have you changed your mind on AGI being a function of compute vs a special alg we haven't found?
1
5