You can add location information to your Tweets, such as your city or precise location, from the web and via third-party applications. You always have the option to delete your Tweet location history. Learn more
1/ On the question of skin tone bias in DeOldify: We take this issue seriously, and have been digging into it. We're going to be overhauling the dataset to make sure it's driving more accurate decisions in the next few weeks. Overall, there seems to be a red bias in everything.pic.twitter.com/VYhS0eC5xS
2/ "Everything" includes ambiguity in general object detection. It can even be seen in Caucasian colorings (see example attached- left is original, right is DeOldify). So overall it seems that DeOldify needs a better dataset than just ImageNet. We're prioritizing it -now-.
3/ That being said, it's a challenging problem and it'll be hard to verify. Everybody is different, so making this work for all cases will probably be something we'll need years to perfect. So please keep an eye out for updates and keep trying it out for yourself!
I wonder if it would be more fruitful to design a mechanism to "tell" the model someone's correct skin tone rather than trying to have it guess everything based on a finite dataset. Maybe the model could "ask" for the appropriate color of regions it's unsure of.
I’d love to make it as automatic as possible but yes, this is probably ultimately necessary to round out the edge cases at least. Now how to do it...not sure on that yet. Sounds very much like exemplar based coloring.
If you ask a skilled Dr., they will tell you that everyone is a different color and that being able to see that has significant diagnostic value. Even diet shifts a person spectral profile noticeably, and painters know there is an interaction with their enviroment's light sphere.
Thank you Jason for your work. Very much appreciated.
There's no "correct way" to decolorize to B&W. So interesting to see if you could find a reasonably correct way to colorize. :)
Interesting! I do know that saturation increases with render factor decreasing.
The dataset I've been using doesn't seem to cause the problem with the red bias, but with any higher RF values it seems to make skin grayish.https://www.colorize.ml/colorized.php?image=db5035be205c66aad4b908e5e641635d711118443b9416e5f8284dcac1210bfe-d1c7b952e1998b6530f947564701e16a.png_15 …
What’s the datase? Can you share?
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.
