Ask HN: Will AI-generated images flooding the web pollute future training data?
7 by albert_e | 1 comments on Hacker News.
We are seeing tons of AI-generated images from Dall-E, StableDiffusion, Midjourney, etc. flooding the internet. This will only increase. Not every such images has a distinct mark that denotes it as AI-generated. They could be mistaken for real photograph or real work of (digital) art by a human. Especially by an algorithm. I also understand a lot of today's cutting-edge models are trained on images scraped from the web. Not sure what curation happens but it cannot be foolproof. Will future AI models that generate "realistic" images feed on this as input and generate images that mimic some of these attributes -- creating some kind of feedback loop that will eco for generations of models? Has anyone already thought of such issues -- not just with images but with AI-generated text, data, music, etc. Curious to know what is the thinking of this group here.
No comments:
Post a Comment