Remastured is an innovative project employing the latest technology to restore some of the oldest - and steamiest - erotic films ever produced. Using machine learning and thousands of adult images from Pornhub.com, we taught AI a dirty new trick: how to colorize vintage porn for a modern audience.
In recent months, incredible restored videos from the early nineteenth century have started popping up online. These videos have been colored and scaled to 4K, thanks to the help of artificial intelligence (AI). Just a few months ago, only big FX studios were capable of producing such amazing works, but now deep-learning algorithms have learned how to do the same thing and the results are equally impressive.
But although these newly restored scenes are quite spectacular from a technical point of view, the subject matter (mostly everyday scenes) leaves much to be desired. So with this powerful technology quite literally at our fingertips, we decided to use those same innovative techniques to pay tribute to porn lovers by remastering vintage porn from a bygone era.
That's why Pornhub, the world's most popular adult website, teamed up with creative studio Espadaysantacruz and the creative agency Officer & Gentleman to create Remastured.
Our mission was to restore films from the very beginnings of erotic and pornographic cinema: From the first kiss, to the first nude, to the first hardcore scenes.
And it turns out that porn's early years - going as far back as 1885 - contain a treasure trove of filthy films, some even shot by the likes of Thomas Edison and Georges Méliès. These highly imaginative works not only have elaborate plots but feature some of the modern porn affcionado's favorite categories: co-eds, masseurs, orgies, threesomes, foursomes, spankings, cosplay, strap-ons, and more. The variety is impressive, but it's the kinky content that will keep you coming back for more.
The Remasturing Process
The process of remastering these films was done by several AI algorithms with virtually no human intervention whatsoever. These are the main steps in the process:
Preparation of images: noise reduction | sharpening | contrasting
Digital remastering: cleaning artefacts | stabilization | flickering reduction.
Colorize using deep learning.
Boosting to 60 fps.
Rescaling to 4K resolution.
The remastering process is a real challenge, as all porn and erotic films from the turn of the century are of very low quality. At the same time, the standard deep-learning models to colorize have been trained with nearly a million SFW images, but we needed our AI to understand nudity.
That's why we decided to retrain the model with thousands of pornographic images to see how our AI learned about the most intimate parts of the human body - nipples, pubic hair, penises and all of the below.
The Machine Learning training
We've based the project on DeOldify.
To put it simply, the training process of this machine-learning algorithm consists of giving it many pairs of black and white images, along with the colorized version. After a while, the algorithm is able to generalize. In other words, it is not only capable of colorizing images that were provided as examples, but it can also colorize images that it has never seen before.
Using thousands of videos of all genres and types from pornhub.com, we extracted more than 100,000 images to use for training. Here is an excerpt of our training set:
This gif contains a 1% of the training dataset and it is really NSFW!!!
After seeing just a few thousand images, the AI was capable of colorizing them. When starting out, the AI was only able to distinguish nudes. The first colorized images were surprisingly interesting:
For the AI's final model training, we started from DeOldify's model "artistic" and we added new information about pornographic nudes.
From left to right: (1) a model with just 7k porn images, (2) the artistic model from DeOldify, (3) the stable model from DeOldify, (4) our final model porn + artistic.
Our final result is not as colorful as the "artistic" model, but it does a very good job with skin tones and we can see more details on the faces.
An interesting experiment consists of recoloring contemporary images where we already know the final result. If we look at the image below, the algorithm does an excellent job with skin tone and hair, for example, but it can't color the underwear properly.
Remasturing Videos
The first step in remastering is to improve the quality of the material as much as possible - stabilising the image, eliminating flickering, and improving contrast and sharpening. We did this using different DaVinci Resolve plugins.
After the manual process we used the project "DeepRemaster: Temporal Source-Reference Attention Networks for Comprehensive Video Enhancement"
(https://github.com/satoshiiizuka/siggraphasia2019_remastering) that does a very good job at cleaning image artefacts.
The next step is to convert the videos to 60fps, giving them fluidity and a more contemporary look. This process is carried out with the DAIN project (Depth-Aware Video Frame Interpolation) (https://github.com/baowenbo/DAIN).
There is also a ready-made application that does not need coding (https://grisk.itch.io/dain-app)
To scale the videos to 4K resolution, we used the Topaz Software. Sometimes the quality of the videos was affected by scaling, so we decided to remaster them to 1080p.
A quick look:
Here we can see a the full movie La Voyeuse from 1924:
or going directly to the pornhub's gallery:
A final word
We created Remastured so that any adult can experiment with this titillating technology and make their own restored adult films. After all, we want all of Pornhub's users to be able to appreciate the history of this much beloved genre, while helping to contribute to it's future.
Client: Pornhub
Agency: Officer&Gentleman
May - 2021
