This Bench Does Not Exist | Hacker News Reader

This Bench Does Not Exist

shkspr.mobi

114 pointsedent50 comments

I'd like to see someone flip this format - This Bench Does Exist - where they physically recreate these slightly deformed benches/whatever. A bit like glitch-inspired art.

I also want to commission a bench labelled

    This Chair Is Left To Rest "IN LOVING MEMORY
     OF WIFE MRS. & DOG SIT AND VISIT MY SPAIN

and just let the future wonder wut.

If you do, please upload a photo to https://openbenches.org :-)

> "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."

That's super overkill. If you just transfer learn from a good checkpoint you'll get good results in a day or two on a V100 (probably even a P100), especially at just 512x512. (I've done this lots of times and a colab pro account for $10 is more than plenty.)

Oh, that's interesting. Do you have a tutorial on how to do that?

I learned everything I know about stylegan (prior to experimenting myself) from this channel: https://www.youtube.com/c/ArtificialImages/videos For stylegan this playlist was useful: https://www.youtube.com/playlist?list=PLWuCzxqIpJs8ViuBIUtAk... (in particular this video https://www.youtube.com/watch?v=KxXqCsl-z2k&list=PLWuCzxqIpJ...). Or just go straight to this colab: https://colab.research.google.com/github/dvschultz/stylegan2...

I'd recommend sticking with stylegan2 for the moment though since it's much easier (cheaper time/computation wise) to transfer from (and the result is about the same - in fact better at the same time and computation budget).

That's amazing - thanks. Gives me something to play with this weekend :-)

what checkpoint would you use though? AFAIK there are no existing datasets trained on benches.

You could just use ffhq/faces, it's fine. You'd think it gets stuck in a local minima but it really doesn't. Transfer learning from any 'finished' checkpoint is much much much faster than doing it from scratch.

Mapillary had object verification tasks for benches in the past, but I'm not sure they're part of their public dataset.

Inquiring minds want to know what would happen if somebody added the Jamaica Pond Taco Bench to the training set...

https://www.atlasobscura.com/places/jamacia-pond-bench

What an incredible bench! In all honesty, I doubt a single "weird" bench would disrupt things too much. I didn't go through every photo in the training set - so undoubtedly a few non-benches slipped through.

Hah! That actually looks like something I wouldn't have been surprised to see already in the output.

It'd actually be pretty funny to include a picture of it in the "full-on psychedelic hellscapes" section and see if anybody notices.

> COPYRIGHT?!?!

> 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.

The generated images are derivative works. If a portion of a source image can be recognized and a case were brought, a court may decide that the generated image infringes on the copyright of the owner of said source image.

For those images made available by the owners under, CC-BY-SA 4.0, remixing is permitted. However, to fulfill the terms of the CC-BY-SA 4.0, you must add explicit attribution, which has not been done.

https://creativecommons.org/licenses/by-sa/4.0/

For those images under CC-BY-SA 4.0, if you 1) add attribution and 2) explicitly license your generated images under CC-BY-SA 4.0 as well, that would fulfill the terms of the license and you would be in the clear. Adding attribution would be laborious, since each image would need to be accompanied by attribution for at least every recognizable source image and arguably every image in the model.

The practical risk is not the same as the reasoned legal exposure, since a copyright holder would have to take action and most image owners will not bother.

Similar analysis applies to all AI-generated creations based on a corpus: they are derivative works of everything in the corpus, and an infringement judgment is possible based on the same criteria that would apply for a manually created collage, copy, pastiche, "tribute", etc. AI is not copyright washing.

Love it! But how soon until spammers start using fake benches to try and get us to buy garden equipment

They'll stop at nothing!

Or using things like https://thispersondoesnotexist.com/ for social media profile pictures

had anyone done a nearest neighbor lookup / cosine similarity from the output of a styleGAN to the training set? I can never quite wrap my head around how much is being hallucinated vs merely regurgitated.

Generative machine learning is still super early. In the short term a lot of the interesting work in machine learning will continue to be focused on large models run on compute clusters (GPT-3, Dall-E, PaLM, RETRO). In the long term I'm excited about bringing all of that local.

A real turning point will be when models are efficient enough or phones are powerful enough to train generative models locally. Combining generative ML with your own local library of images will lead to some interesting applications.

I’m more amazed that their data source OpenBenches exists.

> OpenStreetMap has a list of (nearly) every bench in the world – but they don’t record whether there is an inscription on it. We tried to send a Freedom of Information request to Westminster Council for a list of their benches – but they don’t keep that information. […] So now it is time to crowd-source the data!

https://openbenches.org/

My wife and I created the source - which helps :-)

At the time, OSM didn't record inscriptions. We now have an API so that our benches can be matched to OSM's IDs if they want. But it is a bit complicated because the GPS measures where the photographer is, not where the bench is. And benches do move around quite a bit.

We also syndicate (if that's the right word) some of the images to Wikimedia - mostly of benches dedicated to "notable" people.

A couple of these images pretty accurately represent what it feels like visually before a migraine strikes. Especially those with grass between the legs.

Interestingly (someone told me) that those artifacts and the videos especially can remind one of psychedelic hallucinogens.

There seems to be some shockingly similar transformations, patterns and artifacts with other, related projects as well, including color patterns, movement, things melting or merging into each other etc.

There is a ton of computation happening in the visual cortex. If disturbing it, connecting different things or "seeing behind the curtain" matches some of these visual computing projects then I think we're onto something.

I am getting the sense that over time, we will collectively building in AI a realm populated by the Platonic Ideal of objects in the human world. But since it is in learning models, we will be able to see the zeros and ones, but still won't really have mental access to that realm, only the ability to recognize the particulars that it produces.

Really cool and fun stuff.

Was I the only one that felt uncomfortable when looking at all these animations of unstable objects?

I see how this can be a chinese water torture kind of device [1].

[1] https://en.wikipedia.org/wiki/Chinese_water_torture

There's some sort of "benchmark" pun lying in the open here, unexploited.

I wouldn't bench on that.

The inscriptions are interesting. The one that reads "Dedicated with love to the memory of Alan (Alan) Hutton" reminds me that it has no understanding of what parentheses mean in a name on the inscription.

To be fair, some of the original inscriptions have things like "in memory of David (Dave) Lister". So you can see why the model "thought" that having identical meaning names in that order would be sensible.

Nah, it has absorbed the internet. That's a call-back to Jurassic Park 3.

Is it reasonable to liken such image generation to projection?

Like when they say your brain is constantly projecting its own model of the world for comparison with sensory input. I like that as a model of how our brain works BTW.

Is it just me or do they all look super off? I definitely believe that they do not exist. Them existing would actually be surprising.

Yeah this one is not quite as good as the other "does not exist" sites. A lot of the ones I've seen have a quite realistic looking bench, but with a very obviously not-real backdrop. Strange smears and artifacts are a dead giveaway.

Would Stylegan3 training have worked on Google Colab? Or are the resources just too limited?

That's an interesting question. I think Colab times out after 24 hours - so it would need a bit more hand-holding. Additionally, there's no guarantee of what resources you'll get with them - so the entire process will take a lot longer.

The timeout doesn't matter much, you can periodically save to drive. I spent like a month (subjectively) training gpt2 on Colab by just letting it run for a while and save the weights to drive and then pick up from there next time. I've done the same with GANs, too over a couple of days.

https://svilentodorov.xyz/blog/stylegan-for-evil/

https://svilentodorov.xyz/blog/gpt-15b-chat-finetune/

Does anyone else find these kind of horrifying in a very difficult to explain way?

Yes! There's a sort of "uncanny valley" effect - https://en.wikipedia.org/wiki/Uncanny_valley

Your brain is pattern matching - but it is also noticing the "flaws" in the pattern.

The youngsters might refer to these as "cursed images."

https://en.m.wikipedia.org/wiki/Cursed_image

Looks like generating credible portraits is easier than generating bench pictures.

WHY

FOR SCIENCE!

But, mostly, to understand what is possible. It's easy to read about what other people are doing, but I wanted to try it out myself.

And I figured that no one else had tried it with benches yet.

And isn’t that the spirit of HN in a nutshell?

The "This Person Does Not Exist" was pretty striking, but with all these copycats (a bench now? really?), it's getting a little silly.

As I interpret it, the point is to demonstrate how anyone with a dataset large enough can now do this. Even memorial benches, because that's what the author had.

Yup, that's pretty much it. I had a huge dataset - so wanted to see what interesting uses I could put it to.

We just need a "This (This thing does not exist) Does Not Exist", sort of a meta does not exist, a sort of Y Combinator (natch) of "this thing does not exist".

This particular trend has a long way yet to metastasize. It violates rule 34 at least. The list of "This X does not exist" is counted in the teens and we will see lots more of it.

They are publishing the result of practicing AI

Reply on news.ycombinator.com