A paper by computer scientists Matyas Bohacek and Hany Farid shows that training AI image generators on AI images leads to a deterioration in output quality, akin to inbreeding and collapsing gene pools in species.
Using synthetic data for training can result in nonsensical outputs in text generation, hinting at a similar adverse effect on AI image generator quality.
Developers need to reconsider training data sources to prevent AI image generators from degrading, posing potential increased development costs.
Incorporating content credentials to identify AI-generated images can help developers exclude them from future training sets, maintaining image generator quality.
Collection
[
|
...
]