Language-conditional models can act a bit like decision transformers, in that you can prompt them with a desired level of "reward". E.g., want prettier #dalle creations? "Just ask" by adding "[very]^n beautiful": n=0: "A beautiful painting of a mountain next to a waterfall."
35
220
59
1,266
n=1: "A very beautiful painting of a mountain next to a waterfall."
1
9
141
n=6: "A very very very very very very beautiful painting of a mountain next to a waterfall."
1
15
3
218
n=22: "A very very very very very very very very very very very very very very very very very very very very very very beautiful painting of a mountain next to a waterfall."

Jun 2, 2022 · 2:37 AM UTC

10
58
16
559
Replying to @phillip_isola
Very very cool👍 Which model is this?
Replying to @phillip_isola
I need a new headshot. If I give you a selfie, can #dalle produce a very very very very very very very very very very very very beautiful and also very very very very very very very very very very very very realistic picture of myself? Or will that make it explode?
1
11
Replying to @phillip_isola
Aww c’mon; the poor machine is doing its best; give it a break, now 🥺
1
Replying to @phillip_isola
I'm not sure it understands beauty. It seems to just be weighting the waterfall higher then the mountain, showing more waterfall in the viewpoint the more verys you put in.
Replying to @phillip_isola
this is great haha :D
1
2
Replying to @phillip_isola
And the limit n to infinity gives a Bob Ross painting?
2
Replying to @phillip_isola
This is interesting. I am surprised because I actually agree with it especially with that last one. Interesting. Very interesting. Very Very interesting. If Spock was here, he would say fascinating. But just once.
1
Replying to @phillip_isola
That's very very very very hillarious
1
Replying to @phillip_isola
Wow that is very (x22) beautiful
Replying to @phillip_isola
why did that work I'm so angry
1