DeOldify to make old pictures look new

DeOldify is a set of Python/Jupyter/Machine Learning tools to add color and generally try to rescue aging photographs. Fully Open Source and requiring fairly heavy-weight hardware I thing that it is one to keep an eye on.

Some examples from the project link:
Maria Anderson as the Fairy Fleur de farine and Lyubov Rabtsova as her page in the ballet “Sleeping Beauty” at the Imperial Theater, St. Petersburg, Russia, 1890.


Edinburgh from the sky in the 1920s

There is even a Colab version that lets you try it out without installing any software (you will need a Google Drive account): https://colab.research.google.com/github/jantic/DeOldify/blob/master/DeOldify_colab.ipynb

5 Likes

Wow, what kind of black magic is this? :slight_smile:

3 Likes

No, it isn’t black magic. It’s black and white magic.

5 Likes

Interesting results: a Canadian silver coin, and Ansel Adams’ “Kearsarge Pinnacles”.

image

Generated on the deOldify CoLab notebook (render_factor=42)
https://colab.research.google.com/github/jantic/DeOldify/blob/master/DeOldify_colab.ipynb

1 Like

My Grand-dad in his Uniform:


Produced:

But mysteriously using a colour shot in the original:
Nan%20Grandad%20Claris
Gave mixed results:

Perhaps there is not enough context in the composition. Fares no better with monochromatic (Luminance) input:


the diagonal shadow of the telephone wire on granddad’s forehead is likely causing some discoloration there…
Some tidy-up in GIMP:
34%20AM
and a run thru the deOldifier:

1 Like

It does this to the moon!

1 Like

Similar results to what can done using this website: https://colorize.dev.kaisou.misosi.ru/?lang=en

Colorize

How about people who aren’t Caucasian? Or pictures that aren’t old; e.g., PlayRaw images that are rendered B&W? As there are many ways to go from colour to B&W, it would be interesting to compare the various B&W interpretations of a single source image.

I tried the Black & While image of Mohammed Ali from Wikipedea and got:

so not too bad on that.

Next I took the colour image of Jackie Chan and used a couple of GIMP techniques to convert to monochrome but he came out much more Caucasian.




Just for fun I tried the famous [Marilyn Monroe](By Published by Corpus Christi Caller-Times-photo from Associated Press - Corpus Christi Caller-Times page 20 via en:Newspapers.com, Public Domain, https://commons.wikimedia.org/w/index.php?curid=37860629) shot

2 Likes

I believe it depends on the original machine-learning training set being diverse enough, ethnically speaking. Most of the papers mention this, and indeed deOldify’s paper has non-caucasian examples.

@Steve_Barnes Thanks for providing some examples. Who wouldn’t want to train with Jackie Chan? :stuck_out_tongue: He visited my dad’s nursing home once and they got to shake hands. Now, the nursing home has a large portrait of the man at one of its locations.

1 Like

That website is hosting the so-called Iizuka, et al. 2016 colorize algorithm. iDeepcolor came a year after in the UC Berkeley paper R. Zhang, et al. 2017. deOldify is newer/different in that it is based on a SAGAN (H. Zhang 2018), instead of a GAN like in previous years’ technology.

Thanks @afre - great story. What I like about this ML example is that it has been made so accessible - anybody with a Google Drive account (free) can go to https://colab.research.google.com/github/jantic/DeOldify/blob/master/DeOldify_colab.ipynb and step through the examples without installing a single item. Then, just before executing the last cell upload their own images to the specified directory on their google drive and run it on them. The Google colab service provides all of the computing power. Note that the Co-Lab notebook was contributed by Matt Robinson - great job.

@HIRAM - not only was the original training of a reasonably ethnically diverse set of images of people it also included scenes with no people in - I think that the author, Jason Antic, has done a very good job on this. Also much kudos to the folks at ImageNet for putting together the training data that was used.

1 Like

Photo taken by my grandfather in Guam during WWII.



deOldify does a solid job on CoLab, which gives you a 11GB GPU. I usually find running the default render_factor=42 just about tops it off when processing your images. The default output was a start for some color rotation/de-flourescification in RT5.5, brief touchup in GIMP2.10:
Notice the ox is slobbering on the sugar cane. :yum:

1 Like

Another Guam photo…


IMG_5774
Interesting make up on that sailor…

Sailor bus is pretty creepy.

1 Like

Looks like someone got their brain splattered out in that door window too? Notice the Sailor hat above the blood stain…

That must explain what happened to the two spies who landed on US-occupied Guam in this micro sub.

Yes, some of these colourized images are quite surreal.