4 ms·

Every single game dev on Earth has been thinking that since the very first GANs, but the state of the art when it comes to generative AIs that are actually usef

Every single game dev on Earth has been thinking that since the very first GANs, but the state of the art when it comes to generative AIs that are actually useful for asset generation is still pretty bad. Much like LLMs are still pretty bad at logic. It seems like we're plateauing in that regard... The thing about assets is that they have to be coherent with each other. You can't just generate them one at a time in a bunch of different styles and sizes or they'll look terrible. In the simplest case you want to generate at least an entire spritesheet for a single character or object. And that's still fundamentally impossible with modern generative AIs

lpat 2y ago
Consistency was for sure the biggest issue when I tried this approach last time. Also most of the images were a bit off, like having misaligned limb or some visual glitch. Sometimes fixing those details can be harder than drawing some crappy image from scratch. I really wish there was a tool where I can select the theme, which objects I need and it would spit out some half decently looking art. I even don't care if it was some simplified, cookie-cutter art as long as it was low-effort and consistent. It could be a real time saver for prototyping.
jncfhnb 2y ago
I’ve been exploring this for a while. Icons are pretty solid with flux these days. Previous models struggled with this. Sprite sheets are achievable or near achievable imo. SDXL was maybe good enough as a foundational model if you could leverage IPAdapter or a Lora to maintain character consistency. Although the former was never effectively tuned for styles I wanted and the latter requires some actual artist output still. Flux as a base model is far more consistent and has good enough controlnets to do sprite generation but probably still needs some a proper IPAdapter running to be stable enough for production. Ideally it’s just a productivity tool for an actual artist
ksymph 2y ago
I can't speak for how well it works but PixelLab [0] claims to support animations, rotation and other things that seem impossible. [0] https://www.pixellab.ai https://www.pixellab.ai
kleiba 2y ago
I did, however, specifically use future tense in my post. I agree that we're not there yet but I don't think that we have plateaud. The rapid spurts in GenAI from nothing to what we've got today in such a short amount of time is not sustainable, but we're still seeing improvements all the time. I don't understand the LLM/logic analogy, other than that there's still things about GenAI that are lacking today. Current GenAI is not streamlined for asset generation. But I think there is a huge opportunity to go into that niche and address the problems that currently exist.

lpat 2y ago

Consistency was for sure the biggest issue when I tried this approach last time. Also most of the images were a bit off, like having misaligned limb or some visual glitch. Sometimes fixing those details can be harder than drawing some crappy image from scratch. I really wish there was a tool where I can select the theme, which objects I need and it would spit out some half decently looking art. I even don't care if it was some simplified, cookie-cutter art as long as it was low-effort and consistent. It could be a real time saver for prototyping.

jncfhnb 2y ago

I’ve been exploring this for a while. Icons are pretty solid with flux these days. Previous models struggled with this. Sprite sheets are achievable or near achievable imo. SDXL was maybe good enough as a foundational model if you could leverage IPAdapter or a Lora to maintain character consistency. Although the former was never effectively tuned for styles I wanted and the latter requires some actual artist output still. Flux as a base model is far more consistent and has good enough controlnets to do sprite generation but probably still needs some a proper IPAdapter running to be stable enough for production. Ideally it’s just a productivity tool for an actual artist

ksymph 2y ago

I can't speak for how well it works but PixelLab [0] claims to support animations, rotation and other things that seem impossible. [0] https://www.pixellab.ai https://www.pixellab.ai

kleiba 2y ago

I did, however, specifically use future tense in my post. I agree that we're not there yet but I don't think that we have plateaud. The rapid spurts in GenAI from nothing to what we've got today in such a short amount of time is not sustainable, but we're still seeing improvements all the time. I don't understand the LLM/logic analogy, other than that there's still things about GenAI that are lacking today. Current GenAI is not streamlined for asset generation. But I think there is a huge opportunity to go into that niche and address the problems that currently exist.