How Erasio Works
A high-level look at the technology powering Erasio's AI engine — covering Google Gemini, Nano Banana, and Omni Flash watermark removal across both images and video.
AI-Based Image & Video Processing
Erasio applies advanced computational methods to analyze and reconstruct media data. Using calibrated reference patterns, the engine identifies and removes embedded watermark layers with high precision — across single images and every frame of a video.
- Reference pattern calibration
- Pixel-level reconstruction
- Adaptive processing per image and frame
Advanced Detection & Reconstruction
The watermark detection system locates the mark's footprint by comparing luminance and color channel data. The reconstruction phase restores the underlying original content without visible artifacts.
- Multi-channel analysis
- Artifact-free reconstruction
- Sub-pixel accuracy
Frame-Accurate Video Pipeline
For video, Erasio decodes the clip with the browser's WebCodecs API, removes the watermark from each frame with the same reconstruction engine used for stills, then re-encodes a clean MP4 — copying the original audio track across untouched.
- WebCodecs decode & re-encode
- Per-frame watermark removal
- Original audio preserved
Browser-Native Architecture
All processing runs inside the Chrome Extension using the browser's built-in Canvas and WebAssembly APIs. There is no server, no cloud uploads, and no third-party processing — ever.
- Canvas API rendering
- No network calls for processing
- Zero data exposure
Download Interception
The extension hooks into Gemini's download flow at the browser level, processing images and videos transparently before they reach your file system. The experience is fully seamless.
- Native browser integration
- Pre-download processing
- Transparent to user workflow
Gemini, Nano Banana & Omni Flash Pattern Recognition
Erasio maintains a library of known Gemini watermark signatures, including the Nano Banana compact-overlay pattern. The engine automatically selects the correct removal strategy based on the detected pattern, ensuring accurate reconstruction across all Gemini model outputs.
- Multi-pattern signature library
- Nano Banana compact overlay support
- Automatic pattern detection & selection
What the engine targets — and what it deliberately does not
Google applies two independent signals to media generated by Gemini, AI Studio and Flow. They work differently, they exist for different reasons, and Erasio treats them differently.
The visible watermark is a small sparkle logo composited into the corner of an image, or into every frame of a video, at a known opacity. Because compositing is a deterministic weighted blend, it is invertible: given the blended pixel, the logo colour and the alpha value, the original pixel is recoverable by arithmetic. This is what Erasio inverts, and it is the whole of what the engine does.
SynthID is a separate, invisible provenance signal distributed across the entire pixel grid rather than confined to a corner. It is engineered to survive re-encoding, compression, cropping and format conversion, and it is readable by detection tooling long after any visible logo is gone. Erasio does not target it, does not weaken it, and does not attempt to. Media processed with Erasio still carries SynthID and still identifies as AI-generated to anything that reads for it.
That boundary is a design decision rather than a technical limitation we are working around. The visible mark interferes with legitimate use of media you generated and own — advertising platforms reject watermarked creative, clients notice it in deliverables and previz reels. The provenance signal serves a different purpose entirely, and stripping it is not what this tool is for. The engine is also calibrated specifically to Google's patterns, so it does nothing at all with watermarks belonging to stock libraries, photographers or other creators.
More on this in the FAQ, the guide, and the acceptable use section of our terms.
Note: Erasio processes images and videos entirely within your browser. We do not collect, store, or transmit any media data. Internal implementation details are proprietary. This page describes the technology at a high level only.
All features
Auto-removal, batch, video, multi-pattern detection.
Try the online tool
No install — drop an image or MP4 and download it clean.
Video without quality loss
Why per-frame reverse alpha-blending beats inpainting on video.