AiPixelClearClean a file

How watermark removal works

Updated

The visible sparkle is a known alpha blend. AiPixelClear reverses that blend within its calibrated footprint and checks the restored result before encoding your file.

How does reverse alpha blending work?

A semi-transparent white overlay mixes a known amount of white with the original colour at each marked pixel. The calibrated alpha map records those amounts. Reversing the equation recovers the underlying colour rather than generating a replacement. This operates on decoded working pixels and leaves the rest of the working image intact.

observed = α × overlay + (1 − α) × original

original = (observed − α × overlay) ÷ (1 − α)

Observed pixelsCalibrated alpha mapRecovered pixels

Which profiles are calibrated?

Profile familyMark sizes
Images36 × 36, 48 × 48 and 96 × 96 px profiles
Video sparkleStandard and compact landscape and portrait profiles
Moved complete markGuided template alignment through Help find the mark

What does edge denoising do?

The FDnCNN model can reduce compression ringing around the restored footprint. It uses a small crop for context, with changes confined to the marked area in image output. The model and WASM runtime load when the denoise stage is used. They are served from the same origin as the tool.

How are exports verified?

Choose 720p or 1080p. Images are restored at source dimensions before resizing; video restoration works at the selected size. Image output is decoded again to verify its dimensions. Video checks include frame count, duration, audio presence, copied audio integrity and sampled-frame fidelity. An uncertain restoration follows the recovery path, with no cleaned download presented.

Which browser image limitations apply?

The browser image path produces 8-bit output from 16-bit PNGs and converts CMYK JPEGs to sRGB. PNG preserves transparency. WebP output uses a verified lossless encoder or falls back to PNG. JPEG uses quality 0.92 and offers a lossless PNG alternative. Identifying source metadata is stripped by re-encoding.

How do other methods compare?

MethodWhat happens to pixelsResolutionExisting exportsProcessingCost
Reverse alpha recoveryRestores pixels using the calibrated blend720p or 1080pYes, supported sparkle exportsYour browserFree here
Blur or clone stampSoftens or copies nearby pixelsDepends on export settingsYes, manual editingYour editorDepends on editor
AI inpaintingGenerates replacement contentDepends on export settingsYes, with a suitable editorDepends on providerDepends on provider
Source-product settingControls the mark on future exportsDepends on source productApplies to future exportsSource productDepends on plan and region
Upload-based removerDepends on the selected methodDepends on serviceDepends on supported inputsService serversDepends on service

What is the tool designed for?

AiPixelClear is built for the visible sparkle on supported original exports. Its method is calibrated to that overlay, with every successful output checked before download.

Supported purposeScope boundary
Original Gemini-family sparkle exportsSynthID stays outside this tool’s scope.
Calibrated reverse alpha recoveryOther logos, text, stock marks, people and objects require a different editor.
720p and 1080p exportsResizing does not recover missing source detail.
Media you own or are authorised to editFollow the source platform’s terms and applicable rights.
Read about visible marks and SynthID

How should I prepare my file?

  1. Download the original from the source product.
  2. Keep the complete sparkle and its surrounding corner.
  3. Select the saved export directly from your device.
  4. Review the restored corner at full size before downloading.

Which engine runs here?

The engine is the open-source gemini-watermark-remover, released under the MIT licence. Its calibrated masks trace back to Allen Kuo’s GeminiWatermarkTool. Video decoding and muxing use the unmodified Mediabunny library. Read the full open-source notices.