We forced an AI to look at thousands of photos of memorial benches. Just because. Here are the results.
You can view more at This Bench Does Not Exist.
Here's a video which shows how the process generates images. It gives you an idea of what the model "thinks" a bench is.
You can download a 300MB StyleGAN neural net .PKL for your own use.
What does a bench look like? How can you recognise it? Does the human brain process images of inanimate objects differently than it does with faces and animals? Is there an "uncanny valley" for seats?
5 years ago, my wife and I launched OpenBenches - a site to crowdsource photos, inscriptions, and locations of memorial benches. We have been astonished at how popular it has. People have uploaded over 55,000 photos of about 22,000 benches! That's around 191GB of photos. Nice ☺
I wanted to take all the photos of the benches, and train a computer to produce artificial images of benches. In a similar vein to This [Cat|Person|Paper] Does Not Exist, it's just an interesting way to show off how good (or not) computers are at generating images based off a large dataset.
It is relatively simple to create this. There is a pre-built set of scripts called StyleGAN3. You can follow along with the tutorial provided. I was doing relatively well, until I hit the system requirements:
1–8 high-end NVIDIA GPUs with at least 12 GB of memory. We have done all testing and development using Tesla V100 and A100 GPUs.
Ah. I don't even have one NVIDIA GPU - let alone 12! And the sponsorship for OpenBenches barely covers hosting. Let along buying expensive hardware.
I put the call out on Twitter and, very kindly, Simon Butcher volunteered some hardware. Aren't people on the Internet nice!
The computation for the 512x512px training set was run for 20 days on an IBM AC922 POWER9 server running RHEL7, utilising all four Nvidia Tesla V100 SXM2 16GB GPUs.
The OpenBenches API lets us grab the URls of all images, and metadata about whether they're a photo of the inscription or the bench. There were about 25,000 photos to download and crunch through.
With that many images, it wasn't possible to manually go through to find images which shouldn't have been in the set. That might have led to some weirdness in the generation process.
OK, so here's an odd question. All of the photos on the site have been licenced under Creative Commons. Mostly BY-SA 4.0 but a few imported from Flickr and other sites with different CC licences.
What's the copyright situation of the generated images?
I honestly don't know. If you do, stick a comment in the box below.
There are a number of "glitches" which seem inherent in AI generated images. Some are just weird little graphical quirks.
It's pretty easy to see where the AI has got a little confused. It is totally recognisable as a bench - but just the wrong side of the "uncanny valley" to look real.
Others, however, are full-on psychedelic hellscapes.
Training a text model using GPT-2 was relatively easy. A few years ago I taught an AI to write Shakespearean sonnets. I used the same Google notebook to read in several thousand inscriptions and then generate new ones.
But… we decided not to use them in the end. It's downright creepy generating the name of someone and their date of death. It would be horrible to stumble on the website and find the details of you or a friend listed on there.
As there have already been lots of attempts to automatically generate semantically valid text based on an existing corpus - we dropped it from the site.
Huge thanks to:
- Simon Butcher for training the model on his supercomputer.
- Tom O'Connor for the
- All the hundreds of people who have generously uploaded their photos to OpenBenches.
Here is another video for you to enjoy:
You can view more at ThisBench.DoesNotExi.st.