1/ To hopefully reduce the need for emails on this in the future: Yes, we 100% cherry pick results to post colorizations on Twitter from DeOldify. Furthermore, the research models we have often produce great cherry picked results but aren't reliable enough for general use.
-
Show this thread
-
2/ Quite frankly it has been much easier to produce a model that is capable of creating stunning cherry picked examples than one that reliable produces good examples! That might sound counterintuitive but that's the nature of the beast.
4 replies 1 retweet 25 likesShow this thread -
3/ Out of dealing with this, I've definitely emerged with the mindset that I strongly favor "human in the loop" approaches to "AI art". And there's that last 5% that will probably take 95% of the effort to close the gap on just in terms of "ground truth".
6 replies 2 retweets 33 likesShow this thread -
Replying to @citnaj
This feedback coming from practitioners of in-production AI systems is invaluable. Do you consider this 5% deep technical work (requiring highly trained individuals) ? Is it creative ? Do you expect it to be automated in the future ? I am asking with the "future of work" in mind.
3 replies 0 retweets 0 likes
Perspective: It took 6 weeks to create something that produced cherry pickable results that went viral back in late 2018. This is the first article I saw pop up after I put the code up that made me go "oh shit": https://boingboing.net/2018/11/02/reverse-entropy.html … .
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.