JonathanFly

This is a neat idea, but the execution was kind of minimum-viable-attempt.

You can do this yourself though, just google some famous paintings with X-RAY images and try it. I doubt this is very accurate but it is fun.

I posted a few https://twitter.com/jonathanfly/status/1175966107091574784

> It seems like the input images that you used are paintings.

Yeah, I love trying things that aren't supposed to work! They often do work and surprise you, or fail in interesting ways.

Not bad but monocular depth estimation has gotten pretty good all around. I made these similar images with basically no expertise, no manual mapping, just trying random single image depth projects from github. I kind of just went with whatever I could get running quickly, didn't even evaulate the quality other than cases where it obviously didn't work.

(Sorry for linking a million tweets but I didn't put these on a blog or anything, that's the only place they exist.)

Your version does look better though! But I was impressed I could just run random images through depth estimates and get anything cool like this.

https://twitter.com/jonathanfly/status/1156799136987013120

https://twitter.com/jonathanfly/status/1153383325974896646

https://twitter.com/jonathanfly/status/1154472832249860100

https://twitter.com/jonathanfly/status/1153120643040337925

I would love to your method take on some Escher paintings, can you try it for me?

>The best part was that there was zero quality control, because anyone could edit anyone's map, so you constantly encountered variations and remixes of popular maps. Apart from a handful of very popular main maps, you often had no idea what you were going to get.

Absolutely the best part. It was so easy to modify maps and just try stuff. You could play a brand new game every night. So much creative flourishing, constantly inventing multiple genres. You might even give it credit for AutoChess since the Dota mod was inspired by a similar Pokemon WC3 map.

It looks like the most comprehensive map archive site went down and never came back up -- all those maps will be lost in time.

>It is entirely possible to build a useful AR/MR/VR map of any indoor space using the Apple U1 chip in just a few minutes few minutes.

I'd love to see the fidelity of this. It seems unbelievable...

Imagine a future where 'this abstract was generated by the model' is in the training material for future papers https://twitter.com/jonathanfly/status/1171551688668471297

I read that meaning you can start with the actual pre-trained GPT-2 models but I never got an answer when I specifically asked if that was the case.

I haven't look into at all myself, but he also said:

>We do provide training code that should work out of the box for gpt2 117M/345M

https://twitter.com/TheRealRPuri/status/1161319745259393024

On the NVIDIA GPT- 2 implementation:

>What would be the largest model one could train across 2x 2080Ti?

>~800M gpt2. this is largely due to the memory required to House parameters + optimizer states. If one uses a smaller optimizer than Adam training something larger should be possible. Make sure to turn on activation checkpointing with —checkpoint-activations

https://twitter.com/TheRealRPuri/status/1161322580126470145

I did some tests on this, trying to find footage I thought would be least well suited to interpolation:

https://twitter.com/jonathanfly/status/1156343739738152961

https://www.youtube.com/watch?v=FRHmGEEemoY

I tried to make full cards, text and cart art and card structure, all in the same model but it didn't work:

https://twitter.com/jonathanfly/status/1124918657220534272

This person trained a nice cart art model: https://twitter.com/mkturkcan/status/1121228356899561472

Lastly just an amusing MTG card mashup:

https://twitter.com/jonathanfly/status/1120560055823347712