I've noticed it's worse when repeatedly generating against output. Ex. in the ChatGPT app, I'll ask the agent to use the original source image for refinements (if modifying a non-generated image) rather than just prompting for changes, otherwise the mottling gets bad.