These general techniques are weak against specific detection models that understand how the original image should’ve been generated. Another example is camera noise modeling since the injected noise won’t be consistent with how a phone camera etc actually behaves, or an ML model for image depth might see that it's at the wrong distance somehow.

Scaling the image down seems like one of the best tricks if you can get away with it. Another cheap one would be shifting the image 4 pixels over to break up the DCT artifact grid.