AI Watermark Remover

Experiment with whole-image reconstruction for GPT-image / Nano Banana using local WebGPU or CPU. Invisible watermark removal is unverified. Also repair visible marks and export WebP locally.

Upload · Adjust · Export

Before and after

Keep the picture. Remove the extra mark.

Upload an image or try the example. Processing runs in your browser.

Invisible watermark: unverified

AI Watermark Remover · What this tool does

Experimental whole-image autoencoder reconstruction and latent perturbation for GPT-image / Nano Banana images. Uses local WebGPU when available or CPU, without uploading images. Detail changes and invisible watermark removal remains unverified. Gemini visible sparkle restoration, local inpainting and metadata cleaning are also available.

How to use it

  1. Upload PNG, JPG or WebP and choose the invisible watermark experiment.
  2. Choose GPU / CPU, working resolution and perturbation strength, then reconstruct.
  3. Inspect the device, timing, PSNR and comparison before exporting WebP.

Practical tips

  • Inputs are limited to 25 MB, 12 million pixels and 8192 pixels per side.
  • First reconstruction loads about 32 MB of models and runtime. Export retains original dimensions, not lost detail. Each run starts from the original.
  • Local inpainting loads about 40.8 MB initially. Keep the mask below 25% of the image. Metadata cleaning does not remove pixel watermarks.

Frequently asked questions

Is GPT-image / Nano Banana SynthID removal verified?

No. This is experimental reconstruction without a local SynthID detector or provider verification. PSNR measures pixel differences and is not a watermark score.

Can it use my GPU?

Yes. WebGPU runs on the user device in supported browsers. Automatic mode falls back to CPU when unavailable or when GPU execution fails. No backend or API key is needed.

Creative workspace / ImgHubExplore all tools