Our approach
A clean corner is more than a missing sparkle.
Removing the bright mark is the first step. The texture, color and edges underneath it matter just as much.
Why can a faint outline remain?
A visible watermark blends a light shape into the image below it. A removal method that assumes exactly the right shape and opacity can work well on an untouched export. When those assumptions do not match, a pale patch, a dark edge or part of the star can remain.
Resizing, screenshots and JPEG compression can change the pixels around the mark. Repeatedly processing an already edited image can also make restoration harder. Start with the original downloaded image when you can.
Read the practical guide: why watermark removal leaves traces, and what to try →
One upload, one restoration model
The tool uses a trained restoration model to repair the lower-right corner. You do not need to choose an algorithm or paint over the watermark. Compare the result before downloading, especially on fine lines, text and smooth backgrounds.
Your browser extracts a small corner, sends that corner for processing, and places the result back into the image. The rest of the image is not sent to the model. The downloaded PNG keeps the original pixel dimensions; converting to PNG does not recover detail that was already lost in the source.
Inspect a result from the current model →
What works, and what to check
- PNG, JPG/JPEG, WebP, AVIF, HEIC/HEIF, BMP and still GIF images, up to 20 MB and 40 megapixels. Both image dimensions must be at least 256 pixels.
- The visible Gemini or Nano Banana sparkle must remain in the bottom-right 256 × 256 pixel area.
- You can try a resized, compressed or cropped image, but the result may be less reliable. If cropping or editing has moved the mark outside that corner, this tool cannot reach it.
- It is not designed for arbitrary logos, centered watermarks, text overlays or video.
The tool targets the visible sparkle. It is not designed to remove SynthID or other invisible provenance signals, and a clean-looking result does not mean an image has no invisible watermark.
Technical details and earlier research
The current server model is P20. It processes a native 256 × 256 pixel region without exposing the model weights to the browser. The model runs on a private server; full-image composition happens on your device.
Classic inverse-alpha removal estimates the image beneath an overlay from a watermark template. Learned restoration instead uses a trained model to predict a repair. Neither approach guarantees the exact original pixels beneath a watermark, and a learned repair can soften or alter local details.
Our earlier AuraTuner research article explains why we explored learned restoration and shows the trade-offs between earlier larger and compact models. Its examples and 30-image internal review describe those historical models. They are not scores for P20, a blind benchmark, or a ranking of current competing services.
See image handling and retention for the exact upload and deletion policy.