Yes in fact there is- multiple options! @alexpolymath recently created this slick site that uses DeOldify: https://colorize.cc/ . The only drawback to it is that it's currently GPU limited, which will mean certain renders won't look -as- good. But it's hella convenient.
-
-
Replying to @citnaj @alexpolymath
If you want to take a shot at using a beefier GPU, you can use the Colab notebook: https://colab.research.google.com/github/jantic/DeOldify/blob/master/DeOldify_colab.ipynb … . It's not as user friendly as
@alexpolymath 's website though, as it's Python code that you run in a notebook. But still- a free GPU and hassle free VM comes with it.2 replies 0 retweets 1 like -
Replying to @citnaj @alexpolymath
It's was surprising easy to run the notebook, but the results are kind of greenish :) Any ideas why?pic.twitter.com/ZS6X6BiQZ0
3 replies 0 retweets 0 likes -
Replying to @rani_horev @alexpolymath
Have you tried it playing around with the render_factor variable? Sometime it’s just a matter of hitting the sweet spot with that as to whether or not image elements are interpreted accurately. Generally higher is better but not always
1 reply 0 retweets 0 likes -
Replying to @citnaj @alexpolymath
Nope, but I can try. What's the render factor? (or where can I read about it?) Thanks a lot!
1 reply 0 retweets 0 likes -
Replying to @rani_horev @alexpolymath
In the notebook code. There’s a comment in there that describes what it does and the max value of 42.
1 reply 0 retweets 0 likes -
Replying to @citnaj @alexpolymath
I should've searched there before asking
2 replies 0 retweets 0 likes -
Looks like the sweet spot is different for each image (between 10 to 25). I wonder if we can add a designated module that colorize people in the image, as many times they are the main object in old photos.pic.twitter.com/BwsS9jXOLS
1 reply 0 retweets 0 likes -
Replying to @rani_horev @alexpolymath
Usually with images like this the optimal factor is around 38-42. Have you tried that? On model improvements: They’re coming! Right now I’m focused on trying to make the model better as a whole rather than trying to specialize anything (like people).
2 replies 0 retweets 1 like -
Replying to @citnaj @alexpolymath
You can look at examples of improvements on this front in images I posted a few weeks ago. The Cork, Ireland 1905 crowd photo is a good example. Basically it amounts to bigger model==better. Seems like the same lessons as talked about in the BigGAN paper.
1 reply 0 retweets 0 likes
Note: I haven’t committed the latest models I’ve experimented with to master yet. So you’re still using an older model.
-
-
Replying to @citnaj @alexpolymath
If you need an extra hand, tweet me :)
0 replies 0 retweets 1 likeThanks. Twitter will use this to make your timeline better. UndoUndo
-
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.