BitShovel
中

From model to app · Rediscover existing media

Find a moment without remembering its filename: EmbeddingGemma 2

Search for a turtle eating and find matching moments in a video. Google’s phone demos show how a small model connects text, photos, sound and video for local retrieval.

The official examples are available through Google AI Edge Gallery, after installing the app and preparing its model. This reading follows public demos and documentation, not a BitShovel phone or private-library test.

Google’s demo searches for Turtle eating and returns candidate timestamps above a player and timeline.
Phone region at about 11 seconds in Google’s demo; side margins removed, interface unchanged. The query is Turtle eating; results can be opened for inspection, and scores indicate a match measure. · Open full-size image
In this article4 chapters

中文

Start with what you remember happening

In Google’s phone demo, a query about a turtle eating returns several candidate moments, each with a timestamp. Selecting one opens the corresponding point in the player. Finding material can begin with a remembered event and lead to a moment inside a file.

EmbeddingGemma 2, released on October 6, powers these examples. The demo presents a concrete use: describe a remembered scene in a long recording, then return to the original clip to check it. It demonstrates a retrieval path; accuracy on personal recordings still needs evaluation with those recordings.

Sources and further reading

中文

One model makes different kinds of media comparable

The first EmbeddingGemma focused on text. Version 2 represents text, code, images, audio and video in one shared space. Think of it as producing numbers for comparison: a sentence and a photo can be matched despite their different formats. The full model has 740 million parameters, with weights available in Google’s public repository.

Another demo opens a photo of a cat on a keyboard and offers a similar-image search. An existing picture can therefore become the query. The model produces comparable representations; the app turns matches into grids, timestamps and files that a person can open.

A photo result in Google’s phone demo: a cat lies on a laptop keyboard, with an option to find similar images below.
Phone region at about 21 seconds in Google’s demo, with side margins removed and its English result interface retained. An opened photo can become the starting point for finding similar media. · Open full-size image
Sources and further reading

中文

Local search also needs an app to index the material

Google’s developer guide separates preparation from retrieval. The app builds a local media index, then compares a query with indexed representations. Gallery offers separate media-search and video-moment demos. Preparing the model and index is a distinct stage from subsequent on-device searches.

Our recent SCM reading tackles the same missing-filename problem using CLIP and SigLIP, with separate paths for visible text and dialogue. These official phone examples add another reference. A whole image, a spoken sentence and an action inside a video require different indexing and result interfaces.

Sources and further reading

中文

Open the original after finding a match

A matching score supports retrieval ranking; it is not automatically a probability of being correct. The useful loop is simple: describe what you remember, inspect a few candidates, open the original, then use it in the album, notes or project you are assembling.

What makes this release useful to follow is the visible connection between a model and everyday media retrieval. Chinese queries, easily confused scenes and longer recordings are worthwhile next checks; the demonstrations here do not establish those results.

Sources and further reading