Works on mac os, if setup is correct

This is not an issue. I just wanted to confirm that this does run on mac os using CPU, installed with Anaconda and installing as per the instructions. As I only saw Linux mentioned in the docs I felt it would be good to mention it in case anyone searches for macOS. Testing on Big Sur and will test on Mojave later.

Installing with anaconda and the yaml file produced no errors and was fast. I set a custom env name with -n which was not in the docs, but just how I setup envs here to quickly enable and disable by name. I installed jupyter with pip with the --user flag just in case.

At first I was getting write errors in Jupyter Notebook when using ImageColorizerArtisticTests.ipynb (I placed my images into ./test_images/), because I placed the downloaded models into ./data/models/. Moving the models folder to the project root fixed this.

Just make sure if you’re on mac to set the device id from device.set(device=DeviceId.GPU0) to device.set(device=DeviceId.CPU). I’m running this on an 8 year old MacBook Pro and while a little slow (understandably), it does work and the results are pretty good with the stock models. By slow I mean it might take a few minutes for an image less than 1k in dimensions on either side, which is still very reasonable on old hardware like this.

Going to test on a desktop later and expect it to be much faster. Very cool that this works on mac without much hassle. I’m not a skilled programmer (know enough to get by here and there) and was still able to get this running while having my morning coffee.

Great work and thanks for making this cool tool available. 👍🏻

[jantic]

Great summary- it pretty much jibes with my experience. It really depends on what you’re looking for. and what kind of images you’re dealing with. The source images aren’t available anymore (hosting costs) but you can see what I used on images in this notebook: https://github.com/jantic/DeOldify/blob/master/ImageColorizerArtisticTests.ipynb .

My work ever since this open source release has involved a great deal of effort into making it so that you don’t have to figure out the arbitrary magic number and things just work. I think it’s pretty much at that point now but don’t expect that to make it to open source any time soon. But I can at least let it be known to those developing their own versions that yes- it’s possible to make things fully automatic, but it’s a lot of work.

[KeygenLLC]

The source images aren’t available anymore (hosting costs) but you can see what I used on images in this notebook: https://github.com/jantic/DeOldify/blob/master/ImageColorizerArtisticTests.ipynb .

Flickr, archive.org and Wikimedia Commons are obscenely good resources for getting public domain images to test with. All images are clearly marked and Google Image Search will be much better and faster to search these sites than their built-in search engines. Just append your Google searches with the domain, such as site:wikimedia.org. The only downside to Google Image search is that they removed some of the search tools, so you can’t specifically search for public domain images anymore. Searching the sites I mentioned you will find mostly PD images anyway.

One exception is Flickr. Flickr’s search engine is slower, and can hang up, but it has multiple public domain filters and many galleries are curated by institutions. Millions of images probably.

Note: Searching Flickr images without safe search enabled will possibly give you results that make you wish you had no eyes. You have been warned.