marked images can be excluded from training data in the future, so that it doesn't eat its own output, which is known to cause model quality to degenerate (an i

marked images can be excluded from training data in the future, so that it doesn't eat its own output, which is known to cause model quality to degenerate (an issue likely to plague training LLMs on post-2022 datasets.) as for training the model to add a watermark, that's an interesting idea but I'm not sure it's feasible/not trivial to circumvent.