Comment

Comment on Indie devs have begun adding a no generative AI stamp to their store pages

<- View Parent

mke@programming.dev ⁨4⁩ ⁨days⁩ ago

That’s an interesting enough idea in theory, so here’s my take on it, if you want one.

Yes, it sounds magical, but:

AI sucks at make it more X. It doesn’t understand scary, so you’ll get worse crops of the training data, not meaningful changes.
It’s prohibitively expensive and unfeasible for the majority of consumer hardware.
Even if it gets a thousand times cheaper and better at its job, is GenAI really the best way to do this?
Is it the only one? Are alternatives also built on exploitation? If they aren’t, I think you should reconsider.

source

Sort:hotnew top

Lumiluz@slrpnk.net ⁨3⁩ ⁨days⁩ ago
•Ok, I know the researching ability of people has decreased greatly over the years, but using “knowyourmeme” as a source? Really? • You can now run optimized open source diffusion models on an iPhone, and it’s been possible for years. I use that as an example because yes, there’s models that can easily run on an Nvidia 1060 these days. Those models are more than enough to handle incremental changes to an image in-game • Already has for awhile as demonstrated by it being able to run on an iPhone, but yes, it’s probably the best way to get an uncanny valley effect in certain paintings in a horror game, as the alternatives would be:

spending many hours manually making hundreds of incremental changes to all the paintings yourself (and the will be a limit to how much they warp, and this assumes you have even better art skills)

hiring someone to do what I just mentioned (assumes you have a decent amount of money) and is still limited of course. • I’ll call an open source model exploitation the day someone can accurately generate an exact work it was trained on not within 1, but at least within 10 generations. I have looked into this myself, unlike seemingly most people on the internet. Last I checked, the closest was a 90 something % similarity image after using an algorithm that modified the prompt over time after thousands of generations. I can find this research paper myself if you want, but there may be newer research out there.
source
- mke@programming.dev ⁨3⁩ ⁨days⁩ ago
  You can now run optimized open source diffusion models on an iPhone, and it’s been possible for years.
  
  Games aren’t background processes. Even today, triple-A titles still sometimes come out as unoptimized hot garbage. Do you genuinely think it’s easy to pile a diffusion model on top with negligible effect? Also, will you pack an entire model into your game just for one instance?
  
  I use that as an example because yes, there’s models that can easily run on an Nvidia 1060 these days. Those models are more than enough to handle incremental changes to an image in-game
  
  Look at the share of people using an 1050 or lower card. Or let’s talk about the AMD and Intel issues. These people aren’t an insignificant portion. Hell, nearly 15% don’t even have 16GB of ram.
  
  it’s probably the best way to get an uncanny valley effect in … a horror game, as the alternatives would be:
  
  spending many hours manually making hundreds of incremental changes
  
  hiring someone to do what I just mentioned
  
  What are you talking about? You’re satisfied with a diffusion model’s output, but won’t be with any other method except excruciating manual labor? Your standards are all over the place—or rather, you don’t have any. And let’s keep it real: most won’t give a shit if your game can show them 10 or 100 slightly worse versions of the same image.
  
  Procedural generation has been a thing for decades. Indie devs have been making do with nearly nonexistent art skills and less sophisticated tech for just as long. I feel like you don’t actually care about the problem space, you just want to shove AI into the solution.
  
  I’ll call an open source model exploitation the day someone can accurately generate an exact work it was trained on not within 1, but at least within 10 generations.
  
  Are you referring to the OSAID? The infamously broken definition that exists to serve companies? You don’t understand what exploitation here means. “Can it regurgitate exact training input” is not the only question to ask, and not the bar. Knowing your work was used without consent to train computers to replace people’s livelihoods is extremely violating. Talk to artists.
  
  I know the researching ability of people has decreased greatly over the years, but using “knowyourmeme” as a source? Really?
  
  I tried to use an accessible and easily understandable example. Fuck off. Go do your own “research”, open those beloved diffusion models, make your scary, then scarier images and try asking people what they think of the results. Do it a hundred times, since that’s your only excuse as to why you need AI. No cherry-picking, you won’t be able to choose what your rube goldberg painting will look like on other people’s PCs.
  
  source