Every official Gemini image output size

Why the list is worth having

The Gemini watermark is placed by rule, not by eye. Its size and its distance from the corner follow from the resolution tier the image was generated at. So if you can tell which tier a file came from, you can predict where the mark is before looking at a single pixel, and the position you predict is exact rather than approximate.

Dimensions are the cheapest way to identify the tier. A file that is 2528 by 1696 is a 2k generation at 3:2 and nothing else, because no other tier offers that size. That single match narrows the search from the whole image to a box a few dozen pixels across.

The alternative is reasoning from the aspect ratio alone, which is much weaker. A 3:2 image could be 632 by 424, 1264 by 848, 2528 by 1696 or 5056 by 3392, and those four carry marks of different sizes at different insets. Ratio-only rules get this wrong often enough to matter.

The list below is the catalogue the tool actually matches against. It is compiled from observed Gemini output, and it is kept current as new sizes appear.

Gemini 3.x Image

Four resolution tiers, each offering the same fourteen aspect ratios. Note that the tier names are approximate labels for the long edge rather than exact pixel counts.

0.5k tier

1k tier

16:9 appears twice at this tier. Both 1376 by 768 and 1408 by 768 are produced, and the second one carries a mark of a slightly different size. See the note below.

2k tier

A separate 16:9 size of 2816 by 1536 also exists at this tier. It uses a much larger margin and a newer alpha map, and is treated as its own case.

4k tier

Gemini 2.5 Flash Image

The earlier family offers a single 1k tier and ten ratios rather than fourteen. The extreme panoramic ratios, 1:4, 4:1, 1:8 and 8:1, are not among them.

Note how 2:3 differs between families: 832 by 1248 here against 848 by 1264 on 3.x. The ratio is the same but the pixels are not, which is often enough to tell you which model produced a file.

Two sizes worth knowing about

1408 by 768

This 16:9 size sits alongside the more common 1376 by 768 at the 1k tier, and it carries a mark of 46 pixels rather than the 48 used elsewhere at that tier. It is a small difference, but two pixels is enough to turn an exact inverse into a visible smear, so it is handled as a fixed exception rather than derived from the tier.

2816 by 1536

Also 16:9, also 2k, but generated with a much deeper inset: 192 pixels from each edge rather than 64. It uses an alpha map introduced in May 2026 rather than the original one. When a tool that assumes the standard 2k geometry meets this size, it looks in the wrong place entirely and reports no watermark found.

Checking what you have

On Windows, right click the file, choose Properties and read the dimensions under Details. On macOS, select the file and press Command and I, or open it in Preview and choose Tools then Show Inspector. In any browser, opening the image in its own tab puts the dimensions in the tab title.

If your dimensions appear in a table above, the file is an unmodified Gemini output and the mark is exactly where the catalogue says. If they do not, something has happened to the file since it was generated.

When your size is not on the list

An off-catalogue size is not a dead end, but it does mean the strongest clue is gone. The usual causes, in rough order of how often they turn up:

• A screenshot. Capturing the image from a screen gives you whatever the screen showed, at whatever scaling the display was using. The mark is still there but it has been resampled, which softens its edges and shifts its position.
• A crop. Trimming any edge moves the mark relative to the corner. If the crop took the bottom or right edge, the inset is no longer what the catalogue predicts.
• A resize. Scaling the image scales the mark with it, to a size that matches no known variant.
• A re-export. Saving through another application can change dimensions by a pixel or two, and will almost always re-compress.

In all four cases the fix is the same: go back to the original download from Gemini and run that instead. The anchor search can absorb small offsets, but it works from a prediction, and the prediction comes from the size.