• There is NO official Otland's Discord server and NO official Otland's server list. The Otland's Staff does not manage any Discord server or server list. Moderators or administrator of any Discord server or server lists have NO connection to the Otland's Staff. Do not get scammed!

Enhancing graphics to 64x64 with machine learning

Do you think it an improvement or not?

  • Definitely an improvement

    Votes: 103 96.3%
  • It's worse

    Votes: 2 1.9%
  • No useful difference

    Votes: 2 1.9%

  • Total voters
    107
I could improve the results if I train the neural network with tibia like sprites instead of anime images, but to do that I would need a huge amount of tibia like 64x64 sprites to model the algorithm.

Do anyone know where I can find huge amount of 64x64 sprites?

For example the algorithm from above the author said that trained on 6000 anime images, I would need something like 6000 tibia like 64x64 sprites. Answering my own question I think I can't find that with ease out there haha.
How about using it on tibia's sprites first, and then train a new one with only the 'good' sprites ? Would it make a difference? :D
 
Last edited:
I could improve the results if I train the neural network with tibia like sprites instead of anime images, but to do that I would need a huge amount of tibia like 64x64 sprites to model the algorithm.

Do anyone know where I can find huge amount of 64x64 sprites?

For example the algorithm from above the author said that trained on 6000 anime images, I would need something like 6000 tibia like 64x64 sprites. Answering my own question I think I can't find that with ease out there haha.

http://tochki.su/tibiame/e/Chilong/pc-en

All the Tibia.me (64x64) sprites of creatures, items(equipments), items(stuff) etc
 
[...]Although that testing was done in the entire game window doing in each sprite separately I suspect that would lead to almost the same results.
Do not "suspect". Suspecting is just fancy way of saying: "I'm waffling". Just check it. Take few sprites, use that scaling magic and combine them in any graphical editor.
 
I could improve the results if I train the neural network with tibia like sprites instead of anime images, but to do that I would need a huge amount of tibia like 64x64 sprites to model the algorithm.

Do anyone know where I can find huge amount of 64x64 sprites?

For example the algorithm from above the author said that trained on 6000 anime images, I would need something like 6000 tibia like 64x64 sprites. Answering my own question I think I can't find that with ease out there haha.

How about get flash client spritesheet and use a tool to cut then into 64x64 ? They would give you a huge amount of 64x64 sprites.
 
Wouldn't be easier just to implement scaling scripts in otclient? Just like in sprite based consoles emulators.
 
just looks like its over-sharpened in photoshop, dont know what the machine learning actually does?
 
just looks like its over-sharpened in photoshop, dont know what the machine learning actually does?

The ML is just a way of training the algorithm to convert a 32x32 pixel sprite into 64x64. What you get with good training (providing the algorithm with 64x64 pixel sprites to learn from) is a properly stretched sprite that's much sharper than the original. It simply looks much better.

Looking into this, adding this algorithm to say OTClient wouldn't be feasible without extensive work. For one, running this conversion algorithm requires a Nvidia GPU as a hardware dependency according to the waifu2x github page, so doing it in the client isn't an option unless you have an OpenCL or other alternative.

You'd be better off training the algorithm and doing a one time conversion, saving a new spr file as 64x64 sprites. That'll cause a few problems, particularly file size increase. Map editors will likely work however you'll need both the 32 and 64 spr files, 32 for map editor and 64 for client until the editor supports a 64x64 tile. Plus, as Flatlander said, I do agree that 64x64 pixel would allow for better customization down the road, however it's such a huge task with little reward or demand.
 
The ML is just a way of training the algorithm to convert a 32x32 pixel sprite into 64x64. What you get with good training (providing the algorithm with 64x64 pixel sprites to learn from) is a properly stretched sprite that's much sharper than the original. It simply looks much better.

Looking into this, adding this algorithm to say OTClient wouldn't be feasible without extensive work. For one, running this conversion algorithm requires a Nvidia GPU as a hardware dependency according to the waifu2x github page, so doing it in the client isn't an option unless you have an OpenCL or other alternative.

You'd be better off training the algorithm and doing a one time conversion, saving a new spr file as 64x64 sprites. That'll cause a few problems, particularly file size increase. Map editors will likely work however you'll need both the 32 and 64 spr files, 32 for map editor and 64 for client until the editor supports a 64x64 tile. Plus, as Flatlander said, I do agree that 64x64 pixel would allow for better customization down the road, however it's such a huge task with little reward or demand.
Ahh alright. thanks for clarifying =)
 
Cool idea - I'll implement it. This feature will be soon available in OTClientV8. Preview: http://************/tibiahd.png

tibiahd.png

http://************/tibiahd.png[img]
 
Back
Top