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/ I've been asked quite a bit about why updated video colorization hasn't come out yet. So the short answer is this: A cool tech demo is one thing (videos I've posted on YouTube), but production quality is another beast entirely.
2/ The original DeOldify V1 took 6 weeks to develop (just images), released in November 2018. Cherry picked results were cool, but your average result was quite glitchy and unreliable. Making something that attracted a buyer took 13 months on top of that! Think about that.
3/ Here's the rub too: I really don't think adding more people would help speed this along much. In my mind, this is one of those things that is analogous to "it takes 9 months to make a baby no matter what."
4/ I've believed from the beginning of this project that the real "pay dirt" in deep learning that is still under explored is the discipline of sticking with the problems beyond scoring high on benchmarks and getting impressive cherry picked results.
5/ That is to say, making something real out of the models that can be reliably used by everyday users. This certainly has happened (style transfer and super-resolution are good examples), but not nearly as much has been done in this area as there is potential I think.
6/ Big reason I think is that, no matter your resources, it's a painfully difficult process and it takes a lot of time. Believe me, it's been so very tempting to give up. So video- it'll happen when it's ready. It'll take time. That's pretty much it.
Your work is great! And all good work takes time, so do not give up! Thanks for bringing all that enjoyment through DeOldify. Rooting for the success of the video version!
Thanks!
Hi. I’ve been working on a similar nn to translate digital video (Alexa) to look like 35mm scanned negative film. Was inspired also by your work. I’m at 99% production quality. Breakthrough was to switch from paired to unpaired training, which made video stable
Ps: not sure if unpaired training or cycle consistency loss made video stable as I added both at the same time with a hybrid paired/unpaired training. Cycle consistency might sound counter intuitive for bw>color but might work well. Try it...
Thanks Jason !Some amazing and honest thoughts about developing something that is really good and useful .I was curious about ,were you working full time on this in the initial phase ?
I'd call it 4/5 time. I had a paying job to go to one day a week.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.
