EmbeddingGemma 2 put the expected clip first more often than CLIP in our small test, at the cost of much slower image encoding. It is available as experimental, optional Enhanced Search in B-Roll Archives 0.5.0. The screenshot was captured during development.

OpenAI’s CLIP connects images with text. Google’s EmbeddingGemma 2 supports multimodal embeddings too. For visual video search, we compare your description with sampled frames to rank clips. That’s how you can find video by description even when its filename tells you nothing about the shot.

The coffee shop in the results

Search for “coffee” in the screenshot and Enhanced Search returns The Sensuous Bean storefront near the top. It doesn’t appear among the visible Standard Search results.

Look at the canopy’s lower-left corner: it reads “FINE COFFEES & TEAS.” That small text is a possible clue behind the match. We haven’t isolated which visual features caused the ranking, and this example doesn’t establish that CLIP cannot read signs. It does show a useful difference when looking for coffee-related b-roll footage.

Storefront photo: The Sensuous Bean on Wheree. View the reference photo.

What we measured

The ranking test used 31 sample-library videos and 29 curated queries, counting how often a chosen target clip ranked first. The screenshot illustrates a separate library.

Measurement CLIP / Standard EmbeddingGemma 2 / Enhanced
Expected clip ranked first 22/29 28/29
Short English queries: expected clip first 16/20 19/20
Multilingual queries: expected clip first 2/5 5/5
Mean warm image encoding 9.1 ms/frame 201.5 ms/frame
Total encoding time, 96 frames 3.14 s 20.95 s

The separate native timing benchmark encoded 96 identical cached frames, with Q8 weights, a 140-token image budget and 512-dimensional vectors for EmbeddingGemma 2. Total encoding time sums all 96 timed image calls, including the first cold request. The 22× slowdown compares warm calls only; these measurements exclude video decoding and complete app indexing.

An earlier native full-library Enhanced indexing run took 19 minutes 3 seconds for all 31 sample videos using cached frames. Once frames are indexed, searches reuse their saved vectors.

Tested on an M4 Mac mini with 24 GB RAM running macOS 26.3. This small development test does not establish performance on other Macs or larger footage libraries.

Shameless plug

We built B-Roll Archives for this job: a video organizer for Mac that searches footage where it already lives. Browse contact sheets, search visually and open the original at a matching sampled moment, with indexing and search running locally. Enhanced Search is experimental and off by default in version 0.5.0. You can explore visual search, follow the setup guide, or download B-Roll Archives to try the current workflow.