THE TECHNOLOGY / INDEPENDENT BROWSER DEMO

Try EmbeddingGemma 2 Video Search in Your Browser

A description and a video can share the same mathematical space. That makes it possible to search for a scene without uploading it.

From a remembered scene to a timestamp

FindVideoMoment samples local video at approximately one frame per second, groups up to four frames into a four-second window, and runs the official video processor. It compares each window’s normalized 768-dimensional embedding with your description’s embedding. The highest dot-product scores become likely matching moments.

The model is Google’s EmbeddingGemma 2. We use its text and visual encoders, with audio disabled. We do not average independent image embeddings or substitute cloud inference.

A small input window, a longer video

The model’s context budget applies to each inference call. Streaming four-second windows keeps raw frame memory bounded; it does not make long videos instantaneous. Later scenes cannot be searched until their windows have been indexed.

Runtime

Transformers.js 4.3.1 · WebGPU · q4

Video budget

70 soft tokens · 1fps · 4-second windows

Model revision

daa72c51243991dfcaf9f9137d2c573d8f7790c0

In the previous POC, the 70-token configuration indexed a 14:19 LibreOffice tutorial in 491.90 seconds and a 22:50 Blender tutorial in 641.06 seconds. Those results came from one Apple M5 / 16 GiB environment. They do not establish performance for every browser or video. See full benchmarks and failures.

Google Video Moments Finder vs. FindVideoMoment Browser Demo

Google’s Video Moments Finder runs inside the Google AI Edge Gallery app on Android and iOS, as described in the official Google announcement. FindVideoMoment is an independent browser tool: it searches local video without uploading it and can run on compatible desktop Chromium browsers with WebGPU. It is not developed, operated, certified or endorsed by Google. Both offer on-device search; the distinction here is the app versus browser experience, not a claim of better accuracy or speed.

The similarly named independent commercial website VideoMomentFinder.com is also separate from Google and from this project. We have not run a comparable retrieval-quality test against it and make no superiority claim.

What is Google’s Video Moments Finder?

It is a Google AI Edge Gallery feature for on-device video moment search. Consult the official Gallery project for its current app availability and requirements.

Can I try a video moments finder online without installing an app?

This browser demo can run local visual search in a compatible desktop Chromium browser with WebGPU. It downloads the model after explicit interaction. It is experimental, and compatibility must be checked on your device.

Does FindVideoMoment upload my video?

No video upload is required. The application keeps the file, frames, description and embeddings in your browser. Network access is still needed for the site and model assets; this is not a promise of fully offline operation. Read the privacy boundary.

How does EmbeddingGemma 2 run with WebGPU?

Transformers.js uses ONNX Runtime to execute the pinned q4 text and vision encoders on a compatible GPU. The browser must obtain a GPU adapter and successfully initialize the model. Merely having a WebGPU property is not enough.

What can go wrong?

Visual matches can be ambiguous. In a frozen LibreOffice test, only 7 of 15 targets appeared among the first three candidates. Tiny interface changes and repeated layouts are difficult; spoken words are not transcribed. Similar consecutive windows can repeat in the results. Try a short video and verify the returned scene yourself.