SynthID Image Scrubber (No‑AI)

SynthID Image Scrubber (No-AI) is a browser-only image scrubber inspired by public reverse-engineering research around Google’s SynthID watermark (notably the reverse-SynthID project). Some research pipelines include model-based regeneration stages; this SynthQuery tool intentionally implements only non-AI transforms: no diffusion, no VAE, no learned inference. The intent is pragmatic: apply a small set of fidelity-conscious transformations that can disrupt watermark phase-consensus patterns while keeping the output visually close to the input. Because detectors vary by platform and can change over time, this tool offers presets and transparent “stages applied,” not promises of guaranteed watermark removal.

You can process one image or a batch. Choose a strength preset and export a scrubbed JPEG. The pipeline combines: - Elastic warp (smooth sub-pixel deformation) to fragment global phase coherence - Resize “squeeze” (downsample → upsample) to degrade sub-pixel structure - Subtle color/contrast/saturation nudges plus light luma noise - A multi-pass JPEG re-encode chain (quality ladder) Presets trade off fidelity and disruption. Gentle aims for minimal visible impact, while Aggressive increases deformation and compression steps for stronger disruption.

Implementation details (client-side): - Decode via `createImageBitmap` into an `OffscreenCanvas`. - Generate a smooth random displacement field and warp pixels with bilinear sampling. - Downsample and upsample with high-quality interpolation (squeeze). - Apply lightweight per-pixel adjustments and luma noise. - Re-encode to JPEG multiple times (quality ladder). This tool is not a detector and provides no validity guarantee.

Security and watermarking research can generate deterministic scrubbed variants for A/B comparisons. Publishing workflows can standardize an export step that reduces brittle high-frequency signatures before downstream processing. Content teams can quickly produce alternate derivatives without involving server-side compute or any AI stage.

Choose a preset (Standard is recommended), upload JPG/PNG/WebP images, and run Process (single) or Process all (batch). Download each scrubbed JPEG or bundle results as a ZIP archive. If an image is too large to process in the browser, resize it first and try again.