A deep learning approach to colorizing images, specifically for Pokemon | Hacker News Reader

A deep learning approach to colorizing images, specifically for Pokemon

github.com

63 pointsfloydsoft11 comments

From a deep learning point of view I totally get the appeal of doing this, however...

Surely for the old black and white Pokémon games it would be easier to just rip the assets from the ROM and ask the Pokémon community to re-colour them as a crowd-sourced effort, I'm sure many fans would love to do it.

Would it then be possible to re-inject the assets back into the ROM? (I'm thinking not, as 1-bit images are way smaller than 4-or-8 bit images, and that's not even considering that the original Gameboy ROMs probably don't understand colour images at all...)

What you can do, and many emulators do, is provide a hook to say "when the game draws texture / sprite X, instead draw higher res texture / sprite Y". Obviously, this can be rendered at a much high resolution than the original system. In the case of Gameboy there is no explicit pixel buffer anywhere, so there is no way for the game to even know the emulator has done this.

> Surely for the old black and white Pokémon games it would be easier to just rip the assets from the ROM and ask the Pokémon community to re-colour them as a crowd-sourced effort, I'm sure many fans would love to do it.

And they have, actually: https://www.pokecommunity.com/showthread.php?t=361713 There's a lot of other mono Gameboy game that have been colorized by fans as well, though for the life of me I can't find the collection of them that I'd stumbled upon like 2 years ago.

I find this to be neat demonstration of new AI techniques, and a glimpse of what could be done to enhance our favorite game worlds. How about generating additional world areas based on existing ones? Or making your own DLC with new weapons, characters, quests? Or training on one game and generating content for another, to create a cross-over between mario and megaman?

Original images (https://raw.githubusercontent.com/cameronfabbri/Colorful-Ima...) are upscaled png images with jpg artifacts, seriously!?

They're frames extracted from video as a convenient way to get a large enough training set (50000 images).

It's probably still worth running them all though a de-noise filter first tho.

Any half-decent "deep learning" system should be noise-tolerant.

I wonder if the results would be improved if the inputs were in the original resolution rather than upscaled. Currently there's some smearing between the "pixels" of the art.

Why Diamond though? That one is so different from the gb/gba versions. Prob better result just to stick with 2nd and 3rd gens.

I think this is pretty neat, thanks for posting!

Reply on news.ycombinator.com