But if you look at the image, the image doesn't look like an actual new yorker cartoon; it has the tells of the RL process (notice how it puts random text everywhere for example?)

If you use local models and actually train them on an artist's output, it will look extremely close to that artist's work and will also match the signature.

The image generator shouldn't exhibit the signature problem after it's been reinforcement learning trained.

I think the OpenAI people have trouble with monitoring their model's output because the model has a ton of quirks that shouldn't be there