Crop vs inpainting vs reverse alpha blending

Three different techniques get called "watermark removal," and they are not the same thing. Cropping trims the mark away with real image content attached. Generic inpainting invents new pixels it thinks look plausible. Reverse alpha blending, the method this project uses for documented marks, solves a known compositing equation to recover the pixels that were actually there before the mark was added.

That difference is not a marketing angle. It is the difference between restoring an image and editing one.

Cropping: simple, but it costs the frame

Cropping removes the region containing the mark, corner and all. It is fast, predictable, and needs no model or algorithm — a rectangle selection is the whole tool. The cost is that it removes real image content along with the mark, and it changes the aspect ratio and pixel dimensions. A square Gemini output cropped to hide a bottom-right sparkle is no longer square. A print at a fixed ratio may no longer fit the frame it was ordered for. A video frame cropped to remove a corner mark changes composition for the entire clip, not just one still.

Cropping also does not touch SynthID. SynthID is a signal distributed across the image, not a fixed pixel patch in one corner, and Google describes it as designed to survive cropping, filters, resizing, and lossy compression. Trimming a corner removes a visible logo; it does nothing to an invisible watermark that was never confined to that corner in the first place. Anyone cropping specifically to defeat SynthID is solving a problem that cropping cannot solve — see visible Gemini watermark vs SynthID for the full distinction.

Cropping is the right call when the corner region was disposable anyway — a square export you were always going to trim to 4:5 for a feed, or a frame edge outside your real subject.

Clone stamp and healing tools

A clone stamp or healing brush samples pixels from a nearby, unmarked area and paints them over the watermark by hand. It is a manual version of the same idea as inpainting: cover the mark with plausible-looking texture. Done carefully on a simple background — flat sky, out-of-focus blur, a plain wall — a clone sample from an adjacent patch can be visually indistinguishable from the original. Done on fabric, hair, fine text, or repeating architectural lines, the sampled patch tends to reveal itself as a seam, a repeated tile, or a texture that does not quite continue.

Like inpainting, a clone stamp substitutes a different, real pixel from elsewhere in the frame rather than recovering the original value under the mark. That is a reasonable tradeoff for hand-reviewed retouching, and a poor one for a batch of files nobody is going to inspect individually.

Generic inpainting and generative fill

The common explanation for this method, and it is an accurate one, is: identify the logo, remove it, and fill the edited area with pixels the model considers plausible based on the surrounding content. That is a fair description of what inpainting does, and it is also the method's core limitation stated plainly. Plausible is not the same as accurate. An inpainting model has no access to the pixels that were actually under the mark — only to the pixels around it — so it is always synthesizing an answer, never recovering one.

That has two practical consequences. First, inpainting can hallucinate detail that was never there: a texture that does not match, a line that does not connect, an edge that looks right at a glance and wrong on inspection. Second, inpainting has no principled way to know when to leave an image alone — it fills a watermarked region whether or not filling it is actually the safer outcome.

Generic inpainting earns its place when no known geometry exists for the mark — an arbitrary creator credit, an agency stamp, a stock-photo overlay — where a plausible reconstruction beats a visible logo and pixel-perfect fidelity was never on the table.

Reverse alpha blending: restoring pixels instead of inventing them

When a watermark is composited onto an image with a known alpha, or transparency, map, the visible pixel is a mathematical blend of the original content and the mark:

observed = alpha × logo + (1 − alpha) × original

If the logo's color and the alpha profile at that pixel are known with confidence, the equation inverts cleanly:

original = (observed − alpha × logo) ÷ (1 − alpha)

That is restoration, not invention. Reverse alpha blending does not guess what belongs under the mark — it solves for the value that was mathematically blended away. This is only possible because Gemini, Nano Banana, Veo, and Google Flow watermarks are known quantities: a documented size, a documented anchor position, and a measured alpha profile, catalogued from real generated output rather than assumed from a single sample.

Before1061 × 1008

After1061 × 1008

The geometry behind it, corrected

Reverse alpha blending only works because the geometry is measured, not eyeballed. Resolving 1024×1024 through this project's own catalog function, resolveOfficialGeminiWatermarkConfig() in src/core/geminiSizeCatalog.js, returns a 48×48 mark inset 32px from the right and bottom edges. That is the current answer for 0.5K and 1K Gemini 3.x output alike. Only 2K and 4K output steps up to a 96×96 mark at a 64px margin — resolving 2048×2048 and 4096×4096 confirms it.

There is a confirmed exception inside that rule: a 2816×1536 output uses the larger 96×96 mark at a 192px margin, not 64px, with its own alpha variant. That is why the catalog exists as data rather than a paragraph someone transcribes once and never re-checks. See Gemini watermark sizes for the full measured catalog and supported sizes and formats for what that means for a given file.

Compare the four approaches

Choosing the right method for the job

• A catalogued Gemini, Nano Banana, Veo, or Flow overlay on an unmodified file. Reverse alpha blending is the right tool — try the browser eraser for images or the video workflow for clips.
• The corner was always going to be cropped for a different ratio anyway. Crop it; there is no reconstruction to preserve.
• An unrelated creator, agency, or stock-photo watermark. This project does not cover it — see what this project cannot remove — leaving inpainting or a clone stamp as the honest options.
• Hands-on retouching on a single hero image. A careful clone stamp under human review can outperform an automated pass on simple backgrounds.

Why the engine fails closed instead of guessing

Reverse alpha blending depends on knowing the compositing model, the alpha map, and the edge profile for a candidate mark with real confidence. When an image has been re-cropped, resized, recompressed, or otherwise altered before processing, that confidence can drop below a safe threshold. In that case the engine's answer is to leave the image unchanged rather than force a low-confidence blend that could damage real content.

That is a deliberate design choice, not a missing feature. A skipped edit preserves the source; a forced, wrong edit does not. See why the remover sometimes does nothing for the specific conditions that trigger a skip and what to check first.

Frequently asked questions

What is reverse alpha blending, in plain terms?

When a watermark is composited onto an image with a known alpha, or transparency, map, the visible pixel is a blend of the original content and the mark. Reverse alpha blending inverts that known blend to estimate the original pixel, instead of guessing or painting over the area.

Why not just crop the watermark out?

Cropping removes real image content and changes the aspect ratio and dimensions, which can break a layout, a print size, or a video frame that depends on the original.

Why not always use generic inpainting?

Generic inpainting invents plausible-looking pixels rather than recovering the actual original ones, and it can hallucinate detail that was never there. It also has no built-in reason to leave an image alone when evidence is weak.

When does reverse alpha blending fail closed instead of guessing?

When the compositing model, alpha map, or edge profile for a candidate mark is not confidently known, this project's engine is designed to skip the edit rather than force a low-confidence blend.

Where to try this

Run an actual file through the browser image eraser or the video workflow to see which method your file qualifies for. For the full quality-preserving workflow around a single removal, read how to remove a Gemini watermark without losing quality. If nothing happens on your file, the troubleshooting guide covers why, and how to remove a Gemini watermark online walks the basic flow end to end.