The Use of Neural Net Deep Learning for Icon Generation

[MarkSheppard]

I finally got around to figuring out something that I've always wanted to try ever since the first papers came out about 2015 on the use of Neural Networks for image translation.

For my first attempts, I'm trying to train a neural network to convert GREEN icons to DESERT icons and vice versa; with the following dataset:



I ran CYCLE-GAN for a few hundred epochs and got this basic quality:



You can see how in Epoch 31, the AI is pretty drunk, but by Epoch 431, it's pretty decent.

I stopped there; because the AI went past "good" and became stupid again.

I then went and "tested" my AI against these images which weren't in the training set to see how well the AI would handle it.



And, well....



I guess maybe I need to run it for 5000 epochs?

[DRG]

I suppose you have a good reason for doing this but a conversion from a solid green to a desert tan Icon takes me seconds to do the basics with PSP and a few more to add in details like turret rings

[Karagin]

Still amazes me, I have two icons I tried to make back for the SP2 and they look meh to okay. The stuff you guys do is amazing.

[MarkSheppard]

Originally Posted by DRG

I suppose you have a good reason for doing this but a conversion from a solid green to a desert tan Icon takes me seconds to do the basics with PSP and a few more to add in details like turret rings

Mostly, it's to learn how to use Neural Nets for fun; as there are some ways you could use Neural Nets:

1.) Use existing icons to generate a "made up" icon based on earlier icons -- there have been papers by people who used scans of NES sprites to make entirely made up sprites. (Generate me a random German looking tank!)

2.) Possibly, as Neural Net Research gets better; make it so that a neural network could be trained to draw a sprite from a good top reference or drawing.

[DRG]

Originally Posted by Karagin

Still amazes me, I have two icons I tried to make back for the SP2 and they look meh to okay. The stuff you guys do is amazing.

and a quick camo using the two previous ( 2x normal game size )



All I can say is the learning curve was steep for me as well and *maybe* only one or two of my early examples still exist in the game ( the only one I can think of that is still in use is the Feldwagen in the WW2 German OOB Icon 304.... that is the only one that goes back to the early days PRE SP2WW2 still in use in the game and I could do a better one now but it's "good enough" to be kept it.....just 'cause. That one predates Andy joining the team so early summer 1998.

BUT.... 2 decades of trial and error either builds knowledge and skills or you never get that far. squeezing as much out of <256 colours has been a pet project for years and in that time I have learned what works and what doesn't. It took me years to start using "paste as transparent section" and now it's SOP. It's how I isolate a hull and turret from the background to do a full " repaint" of a solid colour and built up camo patterns

I guess as well it's critical which drawing program you use. I've been using Paint Shop Pro ver 7 since it was released and I am sure there are newer ones out there that can do more but this does what I need it to do

[DRG]

Originally Posted by MarkSheppard

Mostly, it's to learn how to use Neural Nets for fun; as there are some ways you could use Neural Nets:

1.) Use existing icons to generate a "made up" icon based on earlier icons -- there have been papers by people who used scans of NES sprites to make entirely made up sprites. (Generate me a random German looking tank!)

2.) Possibly, as Neural Net Research gets better; make it so that a neural network could be trained to draw a sprite from a good top reference or drawing.

Yeah, I can see this being an interesting project just to see what you can get.

[Karagin]

I use PSP7 myself, still learning new things about it.