Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

Image generation has its own problems with non-cancelling noise.

For example, images are often generated with jpeg artifacts in regions but not globally.

Watermarks are also reproduced.

Some generated images have artifacts from CCD cameras

https://www.eso.org/~ohainaut/ccd/CCD_artifacts.html

Images generated from Google Street View data would likely contain features specific to the cars/cameras used in each country

https://www.geometas.com/metas/categories/google_car/



It seems like such an obvious and surmountable problem though. Indeed since 2020 there are robust approaches to eliminating JPEG artifacts, for example - browse around here - https://openmodeldb.info/.


You're right. In order to be a big problem, the error needs to be non-cancelling and inseparable.

That makes me wonder: if you label good data, and generate data with the good label, how much benefit do you get from also training on okay data?




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: