4 ms·
How are these systems trained to make sure they're not polluted with images created by AI already? I can see how it's semi-easy to have a database of images up
by TacticalCoder 4y ago
How are these systems trained to make sure they're not polluted with images created by AI already?
I can see how it's semi-easy to have a database of images up to a certain date (before AI generating algo existed), but what about if you want to train the AI with modern data?
For example, say if I want to query, in 2035: "Flying hippos in the style of McGrundsBulle" where McGrundsBulle is an artist who only started producing art in 2025... How is the AI then not going to trained with "flying hippos" generated between now and 2035 by AI like DALL-E 2 / midjourney / etc.?
I'd argue it wouldn't make sense to train better AIs using current AI's generated pictures for many of them are the stuff of nightmares and it doesn't seem to be because they've been trained by nightmarish datasets. They're just creepy due to how the current AI works and many details are obviously off, which makes them very creepy to humans.
- ethav1 4y agoDatasets can be manually curated to contain primarily original images if this becomes a real issue. For example, classifiers can predict whether an image is generated or not. You could adapt the process used to create laion-aesthetic[0] to remove generated images. [0]: https://github.com/LAION-AI/laion-datasets/blob/main/laion-aesthetic.md https://github.com/LAION-AI/laion-datasets/blob/main/laion-a...
- coffee_beqn 4y agoAh I kind of want to see the alien horror of AIs teaching each other what art is for multiple generations
- IIAOPSW 4y ago"AI's only know how to copy. They could never invent their own genre. Human creativity is safe for now."
- lancesells 4y agoI don't know why someone isn't doing this already. Output giant datasets of the fantasy AI images and then train only on that.
- ska 4y agoWhy do you think this is likely to end up somewhere interesting?
- lancesells 4y agoNot sure it would but I would be interested to see what would happen. Since the current tools are built on what currently exists in the world of imagery what happens if they are trained on the new mashed up and more fantastical imagery that didn't exist before.
- IIAOPSW 4y agoWhy do you think its inherently a problem if the AI can see itself in the mirror? If anything, those sorts of feedback loops are exactly the missing ingredient to true AI imo.