Synthetic porn is coming – no pun intended

Like many of my posts and analyses this title and the general idea is already a few years old. Now websites like thesenudesdonotexist.com are starting to try to make money with artificially generated nude pictures and I better speed up to publish before reality overtakes me.

In the last few years the generation of fully synthetic but photorealistic pictures has made a lot of progress with techniques like generative adversarial networks (GANs) or variational autoencoders (VAs). GANs are capable of creating realistic high definition pictures of human faces, while VAs encode human faces into vectors that allow the generation and modification of pictures of faces. For example by modifying the correct dimensions of the face vectors the resulting picture might have the face turned to one side or have glasses added or have any other detail or characteristic changed that varies between human faces.

Faces generated from VAs can be modified via vector arithmetic.

These results are especially impressive for pictures of faces and the reason is pretty simple: Compared to how two random pictures from the internet might differ, human faces differ along much fewer dimensions. This low dimensionality allows for the precise modeling with todays compute.

Video generation is on much lower level. Video is of course extremely high dimensional and learning from videos still requires enormous amounts of compute. Actually accurately modeling videos for automatic generation seems pretty far away.

However, porn might be the exception here for at least three reasons:
– Porn should be quite low dimensional compared to … well almost any other type of videos really. With enough compute it is probably already possible to model (and generate) stills from porn videos with impressive accuracy.
– There is no shortage of training data.
– There is a strong incentive to create synthetic porn as soon as it becomes remotely feasible.

I assume at some point somebody is going to scrape the funding together to compute porn-stills embeddings with massive compute and then learn the series of embedding vectors that make up a video using a transformer or LSTM. Finally maybe some GANs to fill in high definition detail.

This would allow for strongly customizable on-demand porn and automate a lot of actors and actresses out of their job.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s