Image: GitHub
Why it mattersA buyer or editor can check a submitted image in the browser. The author's own test caught 65 percent of AI images at 6 percent false positives.
A screenshot of a social post, or a product photo, now needs a check before you trust it. Hikaku is a browser extension that reads that check as you scroll. It flags AI-generated images and video as you browse. The visual check runs on your own machine, and nothing leaves the browser. The code is on GitHub at hikaku-io/hikaku-extension under Apache-2.0. Version 0.1.0 was released on 14 November 2026 as the first public cut. The repository had 394 stars on the next day.
The extension works in two layers. First it reads the file for a provenance tag the AI tool left behind. The README says it parses Content Credentials locally in JPEG, PNG, WebP, MP4, MOV, AVIF and HEIC. It also reads embedded generation parameters from Automatic1111, ComfyUI, InvokeAI and NovelAI. If a tag is there, the extension shows it. The README warns that it does not yet verify the C2PA signature. So a tag can be copied by a bad actor, and the author says so.
If there is no tag, the extension runs a SwinV2 visual model on the device through ONNX Runtime. The model is 400 MB with WebGPU in fp16, or 200 MB on WASM in uint8. The extension skips screenshots, UI chrome and flat graphics before the model runs. A right-click adds a second opinion from a separate 23 MB Community Forensics model.
Numbers from the author's own test
On a historical Reddit sample with weak labels, the visual model caught 65 percent of AI images. The false positive rate was 6 percent, or 2.5 percent when film photography was left out. For AI video the same test caught 22 percent. The README says plainly that a missing flag does not mean human-made. It adds that most platforms strip metadata and that no model catches everything. Store listings, deepfake detection and C2PA signature verification are on the planned list.
The extension needs Firefox 140 or Firefox for Android 142. It also works on Chrome and Edge 116. Flag style can be a border in solid, double or dashed, or a corner dot. Preset and custom colours are both allowed.
A buyer, an editor or a recruiter has to make a quick call on an image. Hikaku puts a signal in the browser rather than a separate tool. The accuracy numbers in the README set the limit of trust. A flag deserves a look. A missing flag still needs judgement from the reader.
Source
The repository is at hikaku-io/hikaku-extension under Apache-2.0.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
