Show this thread
The new "all in one" photo restoration model I'm working on is still early in training, but it's already working quite well to my surprise. Samples:pic.twitter.com/S65pyoqcku
-
New conversation
-
-
-
Yo if you need beta testing, i got some of the worst of the worst photos in need of restoration
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
That's fantastic work Jason. What sort of processing power is involved for all the deep-learning stuff?
-
Training this stuff takes 7-10 days on 4 48GB Nvidia Quadro 8000 RTX GPUs currently. That is to say... a lot!!! Running any particular image- a second, give or take, on a single GPU.
- Show replies
New conversation -
-
-
Amazing work!
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
very impressive!
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Looks great! Did this involve new ideas/approaches or was it a fine-tuned application of stuff you've used in the past? Eg. NoGAN, perceptual loss func.
-
A combination. Exact same architecture and training steps as colorization, but different data. I don't use NoGAN anymore- it's something else now (non-GAN based). The data is where all the new effort went on this one.
End of conversation
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.