Describing text so the model finds it
Locate it and characterize it: 'the white caption text at the bottom of the frame', 'the yellow timestamp in the top-right corner', 'the sale banner text across the middle'. Burned-in text usually sits in a consistent region, which makes it one of the easier removal targets to specify — and finishing the instruction with 'restore the background behind it, keep everything else unchanged' protects the rest of the shot.
Handle one text region per pass. Subtitles at the bottom and a logo bug in the corner are two instructions, run sequentially — each pass stays focused, and you can verify each fix independently.
What to expect from the result
Because the clip is re-rendered, the area behind the text is reconstructed, not revealed — the model synthesizes the wall, sky, or scene it believes continues there. Over flat or softly detailed backgrounds this is usually indistinguishable; over busy texture — or, ironically, over other text, like a caption sitting on a storefront sign — inspect the restored region at full size.
Static text in a fixed position is the most reliable case. Subtitles that change line to line, scrolling tickers, and text that moves with the camera are harder — they can still come out clean, but plan on more than one take. Finish the winning take with the video upscaler before delivery.
The rights question, answered straight
Removing text is legitimate when the footage is yours: your old edits, your dashcam, your archive, or client material you're contracted to clean up. It is not a tool for stripping watermarks, credits, or attribution from other creators' work — that violates our terms and, in most places, the law.
A practical test: if you'd be comfortable explaining to the footage's owner why the text is coming off, you're in the clear. If the point is to hide where the footage came from, don't.