Conversation

Notices

1. Embed this notice
   :blobcathug: (jain@blob.cat)'s status on Saturday, 04-Jan-2025 10:15:09 JST :blobcathug:

   • the marbler 🦊
   @mudkip Wrong. It is generative ai, but the difference is that the model got trained with data of games and with consent
   • Embed this notice
     wizzwizz4 (wizzwizz4@fosstodon.org)'s status on Saturday, 04-Jan-2025 11:30:41 JST wizzwizz4

     in reply to

     • the marbler 🦊

     @Jain @mudkip You're half-right. This model probably can be run as generative AI (by providing noise as input, and calling it repeatedly in a particular way), but this application of it isn't, so it's not properly classified as "generative AI".

     But "generative AI" isn't a useful term, so I don't really care about policing it. I shall instead police the policing of it! (Because that's what we all need: meta-policing.)

   • Embed this notice
     :blobcathug: (jain@blob.cat)'s status on Saturday, 04-Jan-2025 11:30:41 JST :blobcathug:

     in reply to

     • wizzwizz4
     • the marbler 🦊
     @wizzwizz4 @mudkip :blobcatgoogly: Hmm... i disagree... The input is a picture with a lower resolution and the model is built to upscale it by inventing details which do not exist but only to a very limited extent.
     I think you wont be able to get such a model to hallucinate anything meaningful since the model isnt trained for that...
     
     Point is, i would classify the process of inventing details as generative AI, no matter how "good" it works. The difference to a conventional diffusion model is that it doesnt have to invent something from nothing which allows the model to be way smaller and faster... But in the end, both do invent details tho