Unwanted subtitles for east Asian languages

When generating videos where characters talk in east Asian languages like Chinese, Japanese or Korean, there is a 99% chance that the resulting video has unwanted (and completely garbled) subtitles. The problem also appears when using Chinese audio to create lip synced videos. It does not matter if a start image is used or not. The subtitles will appear.

There are a number of suggested fixes out there, for example using NAG, using different samplers, but none of these methods work.

This appears to be a dataset issue with a large amount of videos with burn-in subtitles being inside the training set. If possible this should be reviewed and cleaned up. I would rate this a major issue as currently this otherwise fantastic model is unusable for anything with east Asian speech.