AI generated advertisement signage

[damiante]

Advertising signage for businesses in Japan is of the typically Asian style - crammed in, visually dense and noisy, assaulting to the senses.

In order to replicate this feeling, @Jughents will create sign frame assets designed to contain images. These images will be generated by AI and then used to fill in the sign frames, so that every sign looks different and unique.

Suggestions for image classes:

• Characters representing restaurant types (eg "ラーメン")
• Pictures of food or drinks
• Shopping advertisements
• Anime characters

Important: We shouldn't use logos of konbini because these need to be reserved for mechanical reasons to indicate the presence of an enterable convenience store.

[baely]

@damiante I can do this but will need dimensions,shapes, etc.

It'll be more efficient to generate a whole heap of these upfront, even thousands idc. and randomly select than generate at runtime

[damiante]

@baely how reliably can we generate non-squares? I have asked @Jughents to make some sign placeholders with squares in them that can be filled in. Not sure at this stage if I want them to be stable between map loads. If we bake them into the game doesn't loading time go up a lot? Similar issue to the TTS.

[baely]

@damiante about as reliably as we can generate squares. most models are trained on 1024px images and the overall size is bigger impact as the models awareness of surrounding pixels is relevant.

we can still randomly load at runtime but generate once. i believe it will be easier to define ad hoc jobs that can run for while and just spit out countless assets. then at runtime we can just randomly select from that pool of assets

[damiante]

Ideally we would have a variety of shapes. In particular, single squares, groups of 2x2 squares, single vertical rectangles, and many horizontal rectangles (like a list) are common patterns I recall. Given that the resolution required is very low, can we explore generating at low res vs generating at high res and downscaling?

[baely]

If we bake them into the game doesn't loading time go up a lot?

there's no perfect solution.

1. put them into the game binary and load time is long but gameplay is smooth
2. remote load, load time will be unaffected but transiitions between scenes will take a hit
3. AI generate as required, same as 2 but even slower

[baely]

Ideally we would have a variety of shapes. In particular, single squares, groups of 2x2 squares, single vertical rectangles, and many horizontal rectangles (like a list) are common patterns I recall. Given that the resolution required is very low, can we explore generating at low res vs generating at high res and downscaling?

that's what im exploring atm. i'm trying to find some AI model trained on pixel art. otherwise AI generate -> AI downscale may be a good option too. ideally i would have frames to downsize into. even irregular shapes are fine too, most models support masking