What is the Gemini watermark?
The Gemini watermark is a semi-transparent sparkle logo Google embeds into every image its Gemini model generates. It is applied server-side via alpha compositing at the pixel level — meaning the logo pixels are mathematically blended into the underlying image data rather than overlaid on top, which is why it survives most naive screenshot, crop, and re-encode workflows.
How does Google apply the Gemini watermark to AI-generated images?
Google applies the watermark using the standard alpha-compositing formula: watermarked = α × logo + (1 − α) × original. The alpha map defines how strongly each pixel of the logo overlaps the original. On 1024×1024 output the mark sits roughly 64 pixels from the bottom-right corner; on smaller crops it scales proportionally. The Nano Banana variant follows the same formula but uses a smaller footprint and a lower alpha coefficient — making it visually subtler while remaining mathematically identical to remove.
What does the Gemini watermark look like compared to the Nano Banana variant?
The standard Gemini watermark renders as a 96×96-pixel sparkle on a 1280×720 frame. The Nano Banana variant is roughly half that footprint and sits at a lower alpha value (around 0.35 vs 0.5 for the standard mark), making it harder to spot in low-contrast regions of the image. Both are anchored to the bottom-right corner and both contain the same iconic four-point sparkle shape.
When is it legal and ethical to remove a Gemini watermark?
For personal projects, portfolio images, or commercial work where the user holds rights to the generated content, a watermark-free image is often required by the deliverable spec. Removing the visible mark — whether the standard Gemini sparkle or the Nano Banana variant — restores the image to its generated form for downstream use. Note that Google also embeds invisible SynthID metadata across the pixel grid; SynthID is a separate provenance signal, is not the visible mark you see in the corner, and is not affected by removing the visible watermark.
How does reverse alpha-blending differ from AI inpainting?
Inpainting asks a generative model to invent plausible pixels where the watermark sat — it works but introduces variance, halos, and sometimes a smudge that betrays the cleanup. Reverse alpha-blending is purely arithmetic: knowing the watermark image and its alpha map (both of which are fixed for a given Gemini variant), Erasio computes original = (frame − α × overlay) / (1 − α) and recovers the exact pixels the watermark was hiding. No inference, no halo, no model variance. For a deeper technical walkthrough, see the Erasio /technology page.