AI Inpainting vs Blur: Why Reconstruction Looks Better

The fastest way to "remove" a watermark in most video editors is to drop a blur or pixelate filter over it. It takes seconds and needs no AI model at all. It's also almost always obviously a cover-up, not a removal — anyone watching can tell exactly where the watermark used to be. Understanding why blur looks bad, and what inpainting does differently, makes it easier to judge whether a given result is actually usable.

What a blur filter actually does

A blur or pixelation filter takes the existing pixels inside a region — including the watermark itself — and smears or averages them together. The watermark's shape, color, and edges are still technically present in the frame; they're just scrambled enough that you can't read the logo or text clearly. That's a meaningful difference from removal: the watermark's information hasn't been erased, only degraded. And visually, a soft rectangle or blob sitting on top of otherwise sharp footage draws the eye immediately — it reads as "something was hidden here," which can be more distracting than the original watermark.

What AI inpainting does instead

Inpainting takes a fundamentally different approach: rather than obscuring the pixels under the watermark, it discards them entirely and asks a neural network to predict what should be there based on the surrounding context. The model looks at the texture, color, lighting, and structure around the masked region — sky, skin, fabric, a wall, a repeating pattern — and generates new pixels that plausibly continue that context into the space the watermark used to occupy. Instead of a blurred logo, you get a patch of sky, a continued brick wall, or a hand that looks like it was always there, unmarked.

ClearMark AI runs this with a compact inpainting network (based on the MI-GAN family of models) that executes locally through ONNX Runtime Web, using WebGPU when your browser and hardware support it and falling back to a WebAssembly CPU path otherwise. The model has to be downloaded once — it's a few dozen megabytes — and your browser caches it, so subsequent videos start instantly without re-downloading anything.

Why this is harder for video than for a single photo

Inpainting a single photo is a solved, well-understood problem: mask the region, run the model once, done. Video adds a dimension that breaks the naive approach of just inpainting every frame independently. Because the model has no built-in memory between frames, tiny differences in its prediction from one frame to the next show up as visible flicker in the reconstructed area — an otherwise-still patch of wall that subtly shimmers is often a bigger tell than a blur box would have been. Reducing that flicker requires blending information temporally, using nearby frames to keep the reconstructed region visually stable as playback continues, rather than treating each frame as an unrelated image.

Mask handling matters just as much in video as in photos, arguably more since a small error repeats across every frame. Padding the selected region slightly beyond the visible watermark catches soft edges and faint shadows that a tight crop would miss. Feathering — blending the reconstructed pixels into the original frame with a soft-edged gradient rather than a hard rectangular cut — is what keeps the patch from reading as an obvious geometric shape once the video is playing at normal speed.

How to judge whether a result actually looks good

A few concrete things to check rather than just glancing at a single frame:

• Play the clip at normal speed, not just paused. Flicker and small inconsistencies are far more visible in motion than in a still frame.
• Look at the edges of where the mask was, not just the center. A clean result blends into the surrounding texture; a rough one shows a visible seam or soft halo.
• Check busy backgrounds versus simple ones. Flat surfaces like sky or a plain wall are the easiest case for any inpainting model; detailed, high-frequency textures like crowds, text, or fine patterns are the hardest, and results there are naturally less perfect.
• Compare against a blurred version side by side. Even an imperfect inpainting result usually reads as far less distracting than a blurred rectangle sitting on top of sharp footage.

When blur still makes sense

Blur isn't always the wrong tool. If you need to obscure a moving target for privacy — a license plate, a face, a phone number visible for a split second — a tracked blur can be the more practical and honest choice, since the goal there is visibly redacting information, not making it look like nothing was ever there. Inpainting is the better choice specifically when the goal is a clean, natural-looking frame, such as removing a static watermark or logo from footage you have the rights to edit.

Only remove watermarks from videos you own or have written permission to edit. Open ClearMark AI to see the inpainting approach on your own footage, or read our step-by-step guide to get started.