Why Sora is harder than the rest
Sora 2 moves its badge around the frame. That single detail breaks the usual approach — one box over everywhere it goes — and is why so many tools handle it badly.
Most watermarks sit in a corner and stay there, so one mask covers an entire clip — Veo’s and Kling’s among them. Sora 2 moves its badge around the frame. That single design decision defeats the usual approach and is why so many tools that claim to handle it produce poor results.
It also means the honest answer to “can I remove it cleanly” is: partly, and it depends on your clip.
What we do today
You mark the badge once, on the first frame. From there it is located again on every frame, so the mask follows it rather than covering everywhere it might go. Only the badge gets filled — which is the whole difference between a clean result and a large smear.
The tracking works by matching the badge’s appearance rather than by running a detector. A badge does not rotate or change size, it only moves, and that is a problem arithmetic solves in a couple of milliseconds a frame. Matching happens on a high-pass version of the image, so the semi-transparent badge stays recognisable regardless of what passes behind it.
The search looks near where the badge was a frame ago, which follows a badge that glides across the frame. If it jumps from one spot to another instead, the local search loses it and the whole frame is searched; the mask moves to the new spot only on a strong match, so there is no need to split the clip at each jump.
What remains hard is unchanged: a badge sitting over a face or fine detail fills poorly, and because every frame is reconstructed independently, fast camera movement can leave a faint shimmer.
Getting the best result you can now
Pick your clip carefully. A locked-off or slow shot fills far better than a fast pan, because neighbouring frames agree with each other and the invented patches stay consistent. Steady backgrounds hide reconstruction; busy ones expose it.
Box generously but not wastefully. Follow where the badge actually goes rather than selecting the whole frame, and give it a small margin so no fringe of the logo survives — leftover edges read as a smear, which is more obvious than the badge was.
And consider cropping instead. Sora clips are often destined for a vertical feed, where you are reframing anyway. If the badge spends its time near an edge you were going to lose, cropping gives you untouched pixels for free.
What we leave alone, on purpose
We fill the visible badge and nothing else. We make no attempt to defeat anything built to identify AI video, and have no intention of building one. Making AI video harder to identify is not a product we want to sell.
One thing to know: the clip you download is newly encoded, so the C2PA provenance record inside Sora’s original file is not carried over to it. Your original is never modified. Keep it if you need the record.
Is this allowed?
For clips you generated yourself, ordinarily yes, subject to OpenAI’s own terms — worth reading rather than assuming. Using it to pass generated video off as real footage is a different act, and one our acceptable use policy excludes.
The longer version, with the statutes cited, is in our guide on whether removing a watermark is legal.
Try it
IN SHORT
- Mark the badge once — it is then followed on every frame.
- Steady shots and soft backgrounds fill best; fast pans can shimmer.
- Cropping is often better if the badge sits near an edge you were reframing anyway.
- Your download is a new file without Sora’s C2PA record; keep the original if you need it.
- Sora itself has shut down; this works on clips you saved before it closed.
The Sora watermark remover runs in your browser — the clip is never uploaded, there is no account, and it is free up to 10 seconds at 720p.
Clean a Sora clip
Mark the badge once and it is followed frame by frame. Nothing is uploaded.
THE TOOLS THIS COVERS