Swords artificially generated (GAN)

[tejdi]

I think someone created something similar already but in was more about terrain generation. Applying GANs for map generation could lead to outstanding results, of not only generating terrain but the themes of details, vegetation and house patterns on top of it

Just tell us how much money you want for that.

 

[CipsoftStinks]

Just tell us how much money you want for that.

it's free

 

[Night Wolf]

Just tell us how much money you want for that.

I never tried to apply deep learning for maps, but theoretically it could work. The main obstacle nowadays is having time because I'd need to understand how to port the format of otbm maps into something a network could interpret. Even if we train in simple screenshots of maps, we would need a tool that given a screenshot of the map could identify ids and recreate the map (possibly another ML tool for id detection)

 

[oco]

I never tried to apply deep learning for maps, but theoretically it could work. The main obstacle nowadays is having time because I'd need to understand how to port the format of otbm maps into something a network could interpret. Even if we train in simple screenshots of maps, we would need a tool that given a screenshot of the map could identify ids and recreate the map (possibly another ML tool for id detection)

Maybe this help?

GitHub - Inconcessus/OTMapGen: Uses random noise to generate realistic OTBM terrain with auto-bordering.

Uses random noise to generate realistic OTBM terrain with auto-bordering. - Inconcessus/OTMapGen

github.com

GitHub - Inconcessus/OTBM2JSON: OTBM2JSON - A generic framework for programmatic modification of OTBM files.

OTBM2JSON - A generic framework for programmatic modification of OTBM files. - Inconcessus/OTBM2JSON

github.com

(no idea if it fit for gan, just pointing here)

 

[Otfan125]

I never tried to apply deep learning for maps, but theoretically it could work. The main obstacle nowadays is having time because I'd need to understand how to port the format of otbm maps into something a network could interpret. Even if we train in simple screenshots of maps, we would need a tool that given a screenshot of the map could identify ids and recreate the map (possibly another ML tool for id detection)

Good to know you are still exploring these ideas ; I agree that one of the initial steps in developing a GAN capable of producing otbm maps is to have a solid way to feed it into a model. One idea for producing a training set is to grab a random well developed map ( maybe the CipSoft Tibia RL map replicates in the forums ) and extract every 10x10 area from the map with content, and use that as our training data.
The input to the generator could be a 10x10 parsed otbm map with a type of flooring ( i.e., grass, mud, concrete, etc ) and it's job is to populate it with items, walls, 'stuff' so that it looks correct. In other words, a grass terrain with lava in the middle and a small depot around it would probably fail to pass as a 'real' map. But I guess we can tweak some stuff to make things like that happen.