How prompt-based removal works
Write the instruction the way you'd point the object out on set: 'remove the pedestrian crossing behind the subject', 'remove the red cup on the left of the table, keep everything else unchanged'. Location and color anchor the model on the right thing; the closing clause tells it what not to touch.
Because the clip is regenerated rather than patched, the fill behind the removed object is synthesized — the model reconstructs the wall, street, or tabletop it believes continues there. On typical backgrounds this reads clean; on complex or highly specific backgrounds, inspect the fill closely at full size.
What removals succeed most often
Best odds: objects that are visually distinct from your subject, sit in the background or at the edges of frame, and don't physically overlap the person or product you're keeping. Background walkers, parked cars, wall clutter, and stray props on surfaces are routine removals.
Harder honest cases: objects the subject is holding or touching, large foreground occluders, and shots with fast camera motion where the object crosses the subject. These can still work, but expect variation between runs — generate a few takes and pick the cleanest.
A careful removal workflow
One object per pass. Stacking removals in a single instruction — 'remove the cup and the sign and the guy' — dilutes all three; sequential focused passes are more reliable and let you verify each fix before the next. After the final pass, run the video upscaler so the finished clip ships at full quality.
And the rights note a serious tool owes you: edit footage you own or are licensed to modify. Removing incidental background clutter from your own shoot is what this is for — stripping other people's marks or content from footage you don't have rights to isn't.