Free · private · works offline
You remember what happened, not which file or when. Describe it, and Sceneseek takes you to the exact second. Every match is colour-coded. Every frame stays on your computer.
Windows 10 and 11 · 75 MB installer · free and open source · MIT
Colour is information
Red is a long shot. Green is very likely. The colour runs along one scale from weak to strong, so you can read a page of results at a glance. Drag the dial to see how Sceneseek words it.
The percentage compares a frame with your words against the rest of your footage. It is a guide, not a guarantee.
What it does
Plain English works. Scenes, objects, colours, places and actions are all fair game, and results appear as you type. Or speak it, or drop in a picture.
dog + snow wants both. beach -night leaves night shots out.
A curve shows how strongly every moment of the film matches, drawn above a strip of real frames. Flags mark each take. Click or drag to scrub, or step through takes with the keyboard.
Without a single search, Sceneseek groups your videos into topics such as faces, animals, text on screen, swords and battles. Hover a topic and a frame follows your cursor.
Every moment you keep becomes a numbered cut with in and out timecodes and a note. Export a standard edit list for Premiere Pro, DaVinci Resolve or Final Cut, or take a CSV.
Methodology
Sceneseek uses CLIP, an open model trained on images paired with their captions. It places a picture and a sentence in the same space, so a sentence and the frames that look like it end up close together. No tags, no transcripts.
One frame every second by default. Frames that look almost the same as the last one are skipped, so a static shot costs nothing.
The CLIP image encoder (ViT-B/16, run with ONNX Runtime on your CPU) turns each frame into 512 numbers describing what it shows.
The vectors go into a small database on your disk, about 1 KB per moment. Your video is never copied or changed.
Indexing runs in the background and you can search finished videos meanwhile. It is paid once per video.
“dog + snow” becomes two requests that must both match. “-night” becomes something to steer away from.
The CLIP text encoder turns your sentence into the same kind of 512 numbers. Voice input becomes text first.
Your words are compared with every stored frame in memory. Even tens of hours of footage answer in milliseconds.
Raw similarity is mapped onto a 0 to 100 scale, grouped by video, and drawn as colour-coded frames and a match curve.
Privacy
There is no account, no cloud and no analytics. The AI runs where your videos already are.
Everything else
Results update as you type, and your last searches are remembered.
Every video shows where along its length it matches, so you can jump straight to the best part.
Add a folder once and new videos are picked up automatically. Drag and drop works anywhere.
Save the full-resolution frame of any moment as a picture, ready for a thumbnail or a slide.
Ctrl K jumps to any video, topic, page or command. Every action has a shortcut.
A dark edit-bay theme and a paper theme for daylight. Both scale down to a narrow window.
Download
Run the installer, then add a folder. The first launch downloads the search model once. After that it works fully offline.
Sceneseek is new and not code-signed yet, so Windows SmartScreen may say it does not recognise the app. Choose More info, then Run anyway. You can check the download against the published checksum first.
The SHA-256 checksum is published next to the installer on the release page. In PowerShell, Get-FileHash Sceneseek-Setup.exe should print the same value. A portable zip, with no installer, is on the same page.
The app is plain Python. With Python 3.10 or newer, run pip install -r requirements.txt and python main.py. Full steps are in the README.
Questions
Yes. Sceneseek is free and open source under the MIT licence. There is no account, subscription or ads.
No. Indexing and searching happen on your computer, and the app only answers requests from your own machine. The one download is the search model, the first time you start it. The exception is voice search, which uses your browser’s speech service and may send what you say to it. You can ignore the mic and type.
It matches what things look like: scenes, objects, colours, places and actions. It does not listen to speech, so it cannot search a transcript, and it does not know people by name.
They show how closely a frame matches your words compared with the rest of your footage, not a guarantee. Treat green as “very likely”, amber as “worth a look”, and red as a long shot.
It depends on your processor. On our test PC a 10 minute video took about a minute and a half, so an hour of footage takes around ten minutes. It runs in the background, and you can search finished videos while it works.
MP4, MOV, MKV, AVI, WebM, WMV and others can be indexed and searched. The built-in player plays what your system supports. For anything else, open the file in your default player from the screening room.
The export is a standard CMX3600 edit decision list, which Premiere Pro, DaVinci Resolve and Final Cut Pro can import. Each kept moment becomes a six second cut with your note attached.
Yes, the MIT licence allows commercial use. The AI model is OpenAI’s open CLIP model, also MIT licensed.