Object removal and watermark removal run on the same technology — inpainting, generating replacement pixels from surrounding context. What separates them is that watermarks hold still and objects don’t.
That single difference accounts for most of the difficulty.
Why objects are harder
A watermark has fixed position, fixed shape, fixed opacity. Mark it once and the same region applies for the whole clip.
An object has none of those:
- It moves, so the region has to be tracked frame by frame
- It changes shape as it rotates, or as perspective shifts
- It changes size as it moves toward or away from the camera
- It casts a shadow that also has to go
- It may pass in front of other things, so what’s behind it changes constantly
- It has soft edges — motion blur, depth of field — that don’t stop cleanly
Each of those is a place for the reconstruction to break.
The easy case and the hard case
| Scenario | Difficulty | Result |
|---|---|---|
| Static object, locked-off camera, plain background | Easy | Usually invisible |
| Static object, textured background | Easy | Very good |
| Small distraction in a corner | Easy | Invisible |
| Object with a hard shadow | Moderate | Good if the shadow is included |
| Slow-moving object, simple background | Moderate | Good |
| Person walking through frame | Hard | Variable |
| Object passing in front of a face | Very hard | Usually poor |
| Object in front of readable text | Very hard | Text reconstructs as gibberish |
| Reflective surface showing the object | Very hard | Reflection remains |
The pattern: difficulty tracks how much of the background has to be invented, and how predictable it is.
A bin in the corner of a static shot means reconstructing one patch of pavement that never changes. A person crossing a busy street means reconstructing a different, complex, moving background for every frame they occupy.
The shadow mistake
The single most common reason an otherwise clean removal looks wrong.
You select the object, process it, and the object is gone — along with any credibility, because its shadow is still there. A shadow with nothing casting it reads as wrong instantly, even to someone who couldn’t say why.
Reflections are the same problem in a harder form. If the object appears in a window, a puddle or a polished floor, that reflection needs removing too, and it’s often more awkward than the object.
Treat the shadow as part of the object. Include it in the selection from the start.
The locked-off camera trade
Counterintuitive, and worth knowing before you shoot or choose a clip.
A static camera gives you a consistent background to reconstruct from, which makes the job easier. It also means the patched region sits in exactly the same spot for the entire clip, so any imperfection has the whole duration to be noticed.
Camera motion makes tracking harder, and it hides small errors well — the reconstruction changes constantly, and the eye has less chance to lock onto a flaw.
Neither is strictly better. Static shots are easier to process and less forgiving of mistakes.
Steps
- Assess the object. Static or moving? Shadow? Reflection? This determines how much work it’ll be.
- Start from the highest quality source.
- Select the object plus a margin, including soft edges and motion blur.
- Include the shadow.
- Track across segments if it moves.
- Review full screen in motion. Look for shimmer in the patched area — a sign frames were reconstructed independently rather than consistently.
- Export at maximum bitrate.
What won’t work
- Anything in front of a face — facial structure can’t be inferred from surroundings
- Anything in front of readable text — comes back text-shaped and meaningless
- Objects occupying most of the frame — too little intact context to reconstruct from
- Objects reflected in a mirror or water unless you remove the reflection too
- Very low-resolution footage — not enough surrounding detail to work from
The reliable test: if you can’t confidently say what’s behind the object by looking at the footage, the model can’t either. On video you often get a real advantage here — if the camera moves, earlier or later frames may genuinely reveal what’s behind the object, and better tools use that.
On iPhone
MarkOff uses the same tap-to-select approach for objects as for watermarks: touch what you want gone, the app finds the region boundary, and AI reconstructs the area. Separate removal areas across segments handle objects that move through the frame.
Video is processed on-device, and there’s no blur step in the pipeline — see remove watermark without blur.
For fixed-position marks specifically, remove watermark from video and remove logo from video cover the simpler case.
A note on edited footage
Object removal doesn’t raise the copyright questions watermark removal does — you’re normally editing your own footage rather than stripping someone’s ownership marker.
It’s worth being aware that removing people or objects changes what a video depicts. That’s ordinary editing for personal or creative work. In journalistic, evidential or documentary contexts it’s a different matter, and some platforms now surface provenance metadata that records whether a file was AI-edited. See AI video watermarks explained.