The NotebookLM wordmark is a different problem

Why the usual method does not apply

The Gemini sparkle is removable because two things are known: the logo artwork and its per-pixel alpha map. With both in hand, the compositing equation has one unknown and it can be solved directly.

For the NotebookLM wordmark, neither is published. The obvious response is to recover them from the footage: given enough frames where the same mark sits over different backgrounds, you should be able to fit an alpha and a logo that explain all of them.

That turns out not to work, for a reason specific to where the mark is placed. Measured across roughly 6,700 frames of real exports, the background behind the wordmark is a flat light colour in more than 90 percent of frames. Fitting a per-pixel alpha and logo needs the background to vary underneath the mark, and here it mostly does not. The system is badly conditioned, which is the formal way of saying that many different answers explain the data about equally well and none of them is right.

Fitting one anyway and inverting it leaves clearly visible residue. That was tried before it was abandoned.

Rebuilding the strip instead

What does work exploits the same property that defeated the fit. The mark sits in the bottom right margin, where the background is flat or a smooth gradient. So rather than trying to recover what the mark is, each column of the covered strip is interpolated between the clean pixels directly above and directly below it.

This is a modest idea and it performs far better than the alternatives. Scoring each approach by excess local roughness against the untouched band just above the mark, where zero means indistinguishable from clean background:

The flat fill row is worth reading carefully. Its score is negative, which sounds like an improvement and is not: it means the result is smoother than genuine background. Painting a solid colour over the strip removes the mark and the texture with it, and a band that is too clean is as visible as one that is not clean enough. Column interpolation lands just above zero on flat backgrounds and just above zero on textured ones, which is what you want.

Two layouts, and portrait is not a rotation

The mark is anchored to the bottom right corner and scales with the frame, so in principle one reference layout should cover everything. In practice there are two, because portrait exports are not landscape exports turned on their side.

A 720 by 1280 export draws the mark about 1.45 times the size that the landscape rule, scaled by height, would predict, and sets it further in from the corner. The numbers are not close enough to fudge. Scanning a portrait export across scales and positions puts the mark at 160 by 13 with its corner 30 pixels in from the right and 27 up from the bottom, correlating 0.82 to 0.85 against the template on every sampled frame. The box the landscape rule predicts correlates 0.27 on the same frames, which is indistinguishable from finding nothing.

The consequence was that portrait clips silently reported no mark. The predicted box never cleared the threshold, and it sat just outside the search radius, so widening the search alone would not have found it either.

The two layouts also scale with different dimensions. Landscape exports come at several widths for one height, portrait ones at several heights for one width, so each layout scales with whichever dimension is the stable one for that orientation.

Telling a real mark from a lucky match

Correlation alone is not enough to decide whether a frame carries the mark, and the reason is a general one worth internalising.

A caller that checks one fixed position gets a score that cannot be cherry-picked, so a modest threshold works. A search that tries a few thousand placements and keeps the best one is a different statistical animal: taking the maximum of many samples pushes the score up even when there is nothing to find. Measured over an unmarked export, the best placement scores between 0.19 and 0.23, which clears the fixed-position threshold on almost every frame.

A search using only correlation would therefore find a wordmark in footage that has none. What separates the cases cleanly is peak darkening, because a real mark is far darker than whatever picture content happens to win a search:

So a located match has to clear both a correlation floor of 0.3 and a peak darkening floor of 20. The correlation says the shape is right and the darkening says something is actually there.

The mark is not on every frame

Unlike the Gemini sparkle, which is composited onto every frame of a clip, the NotebookLM wordmark comes and goes. It is drawn over some passages and not others.

So the mark cannot be located once and then removed blindly from every frame. Each frame is checked, and frames without the mark are passed through with their pixels untouched. Running the rebuild over a frame that has no mark would smooth a strip of real picture for nothing.

The footprint that is rebuilt also reaches slightly past the located box: five columns either side and three rows. The box is the size that correlates best with the template, and on portrait exports the word runs about four columns past it while the rows fit, so the horizontal fringe has to reach further than the vertical one.

Where it still falls short

Three cases are known to defeat the current detector, all measured on real exports rather than hypothesised.

• The mark over a dark photo. Where the background luma is around 22, the wordmark is a faint lightening rather than a darkening. Neither polarity of the detector picks it up, so the frame passes through unchanged.
• The mark over fine texture. Column interpolation assumes the background is flat or smoothly varying. Over genuinely detailed texture the rebuild leaves a slightly smoothed band. Better than residue, still not invisible.
• A drawing crossing the mark. When picture content runs through the wordmark, the correlation drops below 0.2 and the frame is treated as unmarked.

Each of these is a case where the tool does nothing rather than doing damage, which is the right failure but not a solved problem. The more general version of that choice is in why a file sometimes comes back unchanged.