I just finished my latest project, building an AI art installation at home, generating 100 % unique artworks on the fly. Just push the button below the screen and another one will be displayed. When the button has been pushed, the old artwork is deleted and can't be retrieved again.
Setup:
* An Nvidia Jetson Xavier NX was used for all logic, machine learning inference, art kiosk GUI etc.
* A StyleGAN was used to generate the artworks, trained on ~5k images of abstract art.
* A passive infrared sensor (SR602) was integrated with the Jetson to reduce screen burn-in. When no movement has been detected around the installation within a pre-defined threshold, the screen shuts off until movement is detected.
* A custom control box was built, encapsulating most of the electronics.
but also planning to use it for neural-net based generative art. I wasn't planning on putting a NX in it though, I was thinking of just keeping the Pi Zero in there and have it do all computations in "the cloud" or on a Nano/Xavier box sitting elsewhere on the same network.
I'm currently working on a 3-panel version of the above:
Curious -- since you're only doing the inference/generation on the frame, and since you're not doing it all the time, did you need a Jetson or would an RPI have sufficed? Did you test inference speeds across different edge compute options?
Yeah I think the way I'd do it (personally) is have an RPi constantly generate new images in the background and cache them until storage is maxed out, then when you hit the button it just fetches the next image from cache.
That would allow the frame to be somewhat lower power and also decrease ventilation requirements -- no fan needed.
I don't think much of the actually produced art, but the fact that you laid out your whole process makes this drool-worthy. Now it is just a challenge as to- can I do better?
Thanks! Depending on how many artworks you've created, it might be difficult to train a GAN network on them (due to overfitting). What you might try is to train one network with a lot of random artworks, then use a Style-transfer network to convert the generated pieces into your style.
this comment seeded an interesting idea! Many artists and photographers want to get into the NFT space but they don't necessarily have experience in digital art creation.
If you could leverage AI to generate digital art based on real artist/photographer inputs, perhaps you could create a nice little marketplace business.. or maybe just a simple AI generator plugin for an existing marketplace..
Hmm, I think The Frame 32" was about $500 and the Nvidia Xavier NX costed as much (including import taxes etc). The other stuff (cable channels, screws, MDF etc) was probably around $50. So a total of around ~$1050.
This says a lot about building MDF housing for a Jetson (BTW why not plywood?), and devotes literally half a dozen pretty generic lines to the actual art generation code. The algorithm is not discussed, and even art examples are not shown.
This is sad, because wooden box building guides are abundant, but art generation guides are less so,
A little white balance work would go a long way I suspect. An OLED panel would really make things pop, but the burn-in issue might be a problem unless you had it rotating very regularly.
Would it have been easier to generate these images on some cloud gpu and stream/send the images to a smart tv? To avoid building and fabricating all the hardware components?
Exactly, a big part of the project was to learn about edge-computing and integrating sensors with the GPIOs. But sure, that would've been possible to do.
Really cool. One piece of feedback: Try to get the white balance to better match the warm wight light of the room. The tv will blend in much much better.
Readit News
Setup:
* An Nvidia Jetson Xavier NX was used for all logic, machine learning inference, art kiosk GUI etc.
* A StyleGAN was used to generate the artworks, trained on ~5k images of abstract art.
* A passive infrared sensor (SR602) was integrated with the Jetson to reduce screen burn-in. When no movement has been detected around the installation within a pre-defined threshold, the screen shuts off until movement is detected.
* A custom control box was built, encapsulating most of the electronics.
https://dheera.net/projects/einkframe/
but also planning to use it for neural-net based generative art. I wasn't planning on putting a NX in it though, I was thinking of just keeping the Pi Zero in there and have it do all computations in "the cloud" or on a Nano/Xavier box sitting elsewhere on the same network.
I'm currently working on a 3-panel version of the above:
https://imgur.com/a/3IfKpb3
I didn't make my own frame though, I designed the dimensions and had it custom-built by a frame company, which was surprisingly affordable.
Was the E-ink screen simple to work with?
Curious -- since you're only doing the inference/generation on the frame, and since you're not doing it all the time, did you need a Jetson or would an RPI have sufficed? Did you test inference speeds across different edge compute options?
That would allow the frame to be somewhat lower power and also decrease ventilation requirements -- no fan needed.
I don't think much of the actually produced art, but the fact that you laid out your whole process makes this drool-worthy. Now it is just a challenge as to- can I do better?
Great job.
This one has how many gazillion photos?
Of course you can! Give it a try and share it afterwards ;)
If you could leverage AI to generate digital art based on real artist/photographer inputs, perhaps you could create a nice little marketplace business.. or maybe just a simple AI generator plugin for an existing marketplace..
Quick question: can you use the Nvidia Jetson Xavier NX to train a model? Or can it only be used for inference?
This is sad, because wooden box building guides are abundant, but art generation guides are less so,
(Covered on HN in 2016: https://news.ycombinator.com/item?id=10900439)
* They are incredibly expensive if you want ~32"
* I wasn't sure that it would've worked together with the Nvidia Xavier NX.
This would make it less tv-isih
1. https://www.claybavor.com/blog/a-canvas-made-of-pixels