Generate a clip with Veo, Sora or Kling and it comes back with a badge in the corner. Getting rid of it sounds like a crop away — and for some clips it is. But video watermarks are a fundamentally different problem from image watermarks, for a reason that’s easy to state and hard to solve: there isn’t one watermark in a video, there are hundreds, and on the newer models they move.
Why one frame is easy and 900 frames are not
A 30-second clip at 30 fps is 900 separate images. Every one carries its own copy of the mark, and every removal has to agree with its neighbours — because human vision is extraordinarily good at spotting a region that behaves differently from the rest of the frame over time. A patch that looks flawless as a still can read as an obvious smudge the moment the video plays.
That temporal constraint is the whole difficulty. It’s also why so many “watermark remover” tools produce a permanently blurred rectangle: blur is the one patch that’s consistent frame to frame without needing to understand the footage underneath.
Two kinds of mark, two different problems
Static marks: reversible, not repaired
A logo composited in the same position with the same opacity on every frame is not really destroyed information — it’s the original pixels blended with a known overlay. If you know the shape and the blend, you can solve for what was underneath and get it back. Nothing is invented, nothing is blurred, and the result is lossless in the region.
This is the good case, and it covers a lot of ground: fixed corner logos, export badges, wordmarks that don’t animate.
Floating marks: reconstruction
The newer generators animate their badge. It drifts, scales, and pulses in opacity — the Veo and Gemini sparkle being the canonical example. Now there is no single overlay to reverse: the mark must be located in every frame, and the region it covered has to be rebuilt from what the surrounding frames know about that part of the scene.
Which introduces the three ways this goes wrong:
- The mask misses the envelope. If you compute the mark’s position on a few frames and apply that box everywhere, the frames where it drifted outside the box keep their watermark — the classic “removed in most of the video” result.
- The mask is too generous. Padding the box to be safe means reconstructing far more of the frame than necessary, and reconstruction is always worse than the original. On low-resolution footage, generous padding is exactly how you get a soft patch.
- Frames disagree. Reconstructing each frame independently gives you 900 plausible patches that flicker against each other. The fill has to be coherent over time, not just per-frame.
What a good pipeline actually does
- Detect. Find the mark — including the faint, low-contrast, pulsing case, which is where naive detection quietly fails and reports a clean video.
- Track the envelope. Follow it across the whole clip and build a mask that covers everywhere it goes, tightly, rather than a fixed box big enough for the worst frame.
- Pick the right treatment. Static and known? Reverse the composite, losslessly. Moving? Send it to temporal inpainting.
- Verify per window. Score whether the mark is actually gone in each part of the clip rather than averaging over the whole thing — an average happily reports success while one section is untouched.
The alternatives, and when they’re actually better
| Approach | Cost | Use when |
|---|---|---|
| Crop the mark out | Lose framing and resolution | The badge is in a corner you didn’t need and the clip is over-resolution |
| Cover it with a graphic | A permanent element in your frame | You were going to put a logo or caption there anyway |
| Generate again without it | Credits, and a different clip | A paid tier of the same generator exports clean |
| Detect, track and inpaint | Compute time | You want this clip, this framing, this resolution |
Cropping deserves more credit than it gets. If the mark sits in a corner of a 16:9 clip you’re about to reframe to 9:16 anyway, the crop was already happening and the problem solves itself for free.
The part worth being straight about
There’s a real line here and it isn’t subtle. Cleaning a generator’s badge off a clip you generated, so it can sit in your edit without an unrelated company’s logo in the corner, is ordinary post-production. Stripping a watermark from someone else’s footage to pass it off as your own is not, and no technical article makes it so. The generator’s own terms of service are a third, separate question from copyright — worth reading for whichever tool you used.
Doing it in Moonshot
The watermark remover runs the pipeline above: detection across the clip, envelope tracking for marks that float, a lossless reversal fast path for the known static logos, and GPU inpainting for everything that moves. It handles the Gemini and Veo sparkle, the Sora badge, and the Kling, Runway, Pika and Luma marks. Upload, wait, download — MP4, MOV and WebM in, clean clip out.
For stills there’s a separate image tool that does the bulk version of the same job.
Quick answers
Why is removing a video watermark harder than an image one?
Because there isn't one watermark, there are hundreds — one per frame — and they move. A static logo can be handled almost losslessly, but a floating, pulsing mark like the Veo sparkle changes position, size and opacity frame to frame, so anything that misses it for even a few frames leaves a visible flicker.
Can you remove a watermark without blurring the video?
For a static, known logo, yes — if you know exactly how the mark was composited you can reverse that operation and get the original pixels back rather than inventing new ones. For a moving mark over moving footage there's no original to recover, so the region has to be reconstructed from neighbouring frames, and the quality question becomes how good that reconstruction is.
Does removing the visible logo remove the invisible watermark too?
No. Visible logos, invisible pixel-level signals like SynthID, and provenance metadata such as C2PA are three separate layers living in three different places. Erasing the corner logo does nothing to the other two — a distinction we cover in detail in the three-layers article.
Which AI video watermarks can be removed?
The common ones: the Gemini and Veo sparkle, the Sora badge and wordmark, and the corner logos from Kling, Runway, Pika and Luma. Static corner marks are the easy case; the floating, pulsing ones are the hard case and need frame-by-frame tracking.
Is it legal to remove a watermark?
It depends entirely on whose video it is. Removing a generator's badge from a clip you generated yourself is a different question from stripping a mark off someone else's work, and platform terms of service are separate again from copyright law. Treat it as: your own generations, fine to clean up; anyone else's, don't.
Clean the badge off your own generations
Detection, frame-by-frame tracking and inpainting for Gemini, Veo, Sora, Kling, Runway, Pika and Luma marks — including the floating sparkle that fixed-box tools miss.
Open the watermark remover