# Video Search > Describe what is on screen and jump to that moment in your video. An AI model indexes a frame every 5 seconds on your GPU, and the video is never uploaded. - URL: https://bytescope.dev/en/video-search/ - Locale: en · Also: [ja](https://bytescope.dev/ja/video-search.md) · [zh-Hant](https://bytescope.dev/zh-Hant/video-search.md) - Runs: Your files never leave your browser — all processing is local. - Requires: Chrome or Edge on desktop with WebGPU. Without WebGPU the search cannot run and the page says so; Safari, Firefox and mobile browsers are untested. The first use downloads about 344 MB of model files once. - Updated: 2026-09-08 ## What it does Drop a video and describe what you are looking for: "a dog runs across the lawn", "the slide with the bar chart", "sunset over the sea". You can also drop, paste or choose a screenshot to find the moments that look like it. The page first takes one frame every 5 seconds, turns each frame into a vector with the EmbeddingGemma 2 model on your graphics card, and turns your words or picture into a vector the same way. Search opens as soon as the first frames are indexed. A second pass then fills in one frame per second in the background, and you can pause and resume it. The closest moments come back as up to 12 results, each with a thumbnail and a timecode. Clicking one starts the player there, and "Trim in Video Trimmer" opens that range in Video Trimmer once you choose the same file again there. ### What it needs The first use downloads about 344 MB of model files once; after that they stay in your browser and load from there. The model runs through WebGPU, so you need Chrome or Edge on a desktop computer with a graphics card. Without WebGPU there is no fallback, because the model cannot run on the processor here, and the page says so instead of trying. Safari and Firefox have not been tested, so they are not claimed. The video is never uploaded: frames are decoded and indexed in this tab, and nothing about them leaves your computer. ### Honest limits Frames are ranked by how close they are in meaning to your words; the model does not really understand the video. Concrete things you can see, such as an object, a place, a colour or a slide, work best, and abstract words like "the funny part" work poorly. It only looks at pictures and cannot hear, so a search for "someone singing" can only go by what is on screen; to search what people say, use Voice Notes. The first pass samples one frame every 5 seconds, so until the fine pass reaches a part of the video, a shot shorter than that can fall between two samples and be missed. Indexing time depends on your graphics card: on a GTX 1060 the first pass takes about 3 seconds per minute of video, so a 40-minute film is searchable in about 2 minutes, and the fine pass takes about 15 seconds per minute of video. The vectors are kept in this browser, so if you close the page and come back, choosing the same file picks up where indexing stopped. It opens the same formats as Video Trimmer, MP4, MOV, WebM and MKV that your browser can decode; an HEVC video on a computer without an HEVC decoder cannot be read. ## Limits Does not upload the video, does not listen to the sound (it searches pictures only), does not look at every frame (one frame every 5 seconds, so shots shorter than that can be missed), does not understand the video (it ranks frames by similarity to your words), and does not keep the index after you reload the page. ## FAQ ### Is my video uploaded? No. The video is read from your disk, its frames are decoded in a background worker in this tab, and the AI model runs on your own graphics card. The only download is the model itself, once; neither the video nor your search words are sent anywhere. ### Why does it not work in Firefox, Safari or on my phone? The model needs WebGPU, and it has only been tested in Chrome and Edge on desktop computers. Without WebGPU it cannot run here at all, so the page shows a support table instead. Safari, Firefox and mobile browsers have not been tested, so they are not claimed even where they offer WebGPU. ### How long does indexing take? It depends on your graphics card. On a GTX 1060 the first pass, one frame every 5 seconds, takes about 3 seconds per minute of video, so a 40-minute video is searchable in about 2 minutes, plus the one-time 344 MB model download on first use. The fine pass, one frame per second, then runs in the background at about 15 seconds per minute of video; you can pause and resume it, and search keeps working meanwhile. While the first pass is running you can cancel it and still search the part that is done; the model download and the opening of the file cannot be cancelled. If you close the page, choosing the same file again later continues from the frames already indexed. ### Can it find what people say in the video? No. It only looks at the pictures and never listens to the sound, so speech, music and sound effects are invisible to it. To search spoken words, transcribe the audio with Voice Notes and search the transcript there. ### Why did it miss a short scene? The first pass looks at one frame every 5 seconds, so a shot shorter than that can fall between two samples until the fine pass, one frame per second, has reached that part of the video. Even then a shot shorter than a second can be missed. Each result starts at a sampled frame, so the moment you want may begin a little before or after the timecode shown. ### How accurate is it? In our test on three open animated films with 30 descriptions of what was on screen, the right moment was in the top 3 results for 29 of them. That is a small test: live-action footage and screen recordings were not measured, and abstract or very specific descriptions do worse than plain descriptions of what you can see. ## Related tools - [Video Trimmer](https://bytescope.dev/en/video-trim.md) — Found the moment? Cut that part out without re-encoding. - [Screen Recorder](https://bytescope.dev/en/screen-recorder.md) — Record your screen, then search the recording for a slide or a step. - [Video Info Inspector](https://bytescope.dev/en/video-info.md) — See the codec and container when a video will not open.