00Video analysis in your browser
Drop a video.
See everything in it.
Frame Recon scans a video on your own device. It marks the scenes, follows the objects it recognises and writes down what was said. Then it answers questions about any moment, right down to the small stuff in the background.
- Detector
- 365 classes
- Sampling
- up to 4 fps
- Runs on
- your device
- Upload
- none
- Nothing to upload
- Your video is never stored
- No account needed to scan
01How it works
Three steps. No setup.
- 01
Drop it in
Pick a video file or paste a YouTube link. Files open straight from your device, so there is nothing to upload and nothing to wait for.
- 02
Watch it scan
Models run in your browser. They mark scene changes, detect and follow objects through each shot, and write down what was said.
- 03
Ask anything
Type a question. Frame Recon goes back to the right frames, zooms in if it needs to, and answers with the exact moment.
02Why the answers hold up
It looks again before it answers.
Most tools describe a video once, then answer from those notes. Anything nobody thought to write down is gone.
Frame Recon answers from the picture. Ask about a tiny red hydrant in the background and it finds the moment, pulls that frame from the video, zooms in on the original pixels and checks. If it cannot find something, it tells you where it looked.
- 01Question
What colour is the hydrant behind the far goal?
- 02Locate
Searches the scan, then skims the video for the right moment.
- 03Zoom
Pulls that frame from the video and crops in on the detail at its original sharpness.
- 04Answer
Red, with a darker cap. It is in shot from 0:38 to 0:47.
03What you get
Hours of scrubbing, done in minutes.
Object tracking
365 kinds of object, followed through each shot. See when each one appears and for how long.
Ask the video
Answers come from the actual frames, with timestamps you can click to jump there.
Find anything
Search for an object, a word on screen or something that was said, and go straight to it.
Transcript
Speech to text on your own device. Click any line to hear it again.
Deep scan
One extra pass that writes down everything in each scene, including background objects and text on screen.
Scenes and timeline
Scene changes marked, with a timeline of what is on screen and when.
04The facts
What it does. What it does not.
No big claims here, only what the system does today. Where the answer depends on your computer, that is said too, and the panel marked This device checks yours.
- Opens files straight from your device. MP4, MOV, WebM and MKV. Nothing is uploaded, and we set no size limit.
- Finds 365 kinds of everyday object. It puts a box round each one and follows it from frame to frame.
- Answers questions by looking again. The AI model reads the scan, asks your browser for the exact frames it wants, zooms in on the original pixels and answers with the moment.
- Writes a transcript on your device. Speech becomes text with a time on every line. The language is worked out from the sound, or you can pick one of 99.
- Reads public YouTube videos from a link. A summary, chapters, key moments, and questions about both picture and sound.
- Handles long videos and 4K. The file is read a piece at a time, so a two-hour film needs about the same memory as a short clip.
Show 8 moreShow fewer
- Looks closer for small things. The first frame of every scene is also checked in overlapping close-ups: 4 for HD video, 9 for 4K.
- Marks scene changes. It builds a timeline of what is on screen and when.
- Checks every transcript line for a voice. Speech models can write made-up lines over music and noise, so a line with no voice behind it is marked unclear and folded away, not passed off as something that was said.
- Picks up very quiet speech. A soundtrack recorded far too low is made louder before it is transcribed. So is a quiet voice that comes after a loud one, or after loud music.
- Searches everything it found. Objects, spoken words and deep scan notes, in any alphabet.
- Deep scan writes down what is in a scene. Background objects, text on screen and what is happening. One frame per scene, up to 48 frames a run.
- Tracks objects in a YouTube video. A live scan watches the tab while it plays.
- Saves its progress as it goes. Refresh or close the tab and the scan carries on from where it stopped.
- It does not look at every frame. The scan looks at 1 to 4 frames a second, so something on screen for a fraction of a second can be missed by the boxes. A question can still go to any frame.
- It only boxes the 365 kinds of object it knows. Anything else is found by asking or by a deep scan, not boxed automatically.
- It does not recognise faces. It will not tell you who someone is.
- It can be wrong. Every box has a confidence score and every answer points to a moment, so you can check.
- The transcript is not equally good in every language. On our test clips, in the full scan, it got about 1 word in 20 wrong in English and German, 1 in 15 in French, Spanish, Portuguese and Polish, 1 in 10 in Russian, Turkish and Korean, and between 1 in 5 and 1 in 3 in Arabic, Hindi and Swahili. It does not say who is speaking.
Show 7 moreShow fewer
- The voice check is not perfect. A murmur too faint to make out still counts as a voice, and one short word under loud noise can be marked unclear by mistake. Marked lines are never deleted. They are one click away.
- One language per video. If the speakers switch language part way through, the second language comes out wrong.
- For video files it listens for speech only. Music and sound effects are not analysed. YouTube links are different: the model hears the whole soundtrack.
- It cannot read private, age-restricted or members-only YouTube videos. It does not take links from Vimeo or other sites yet. Download the video and drop the file in.
- It cannot open copy-protected streams. Netflix and similar services will not play into it.
- It is not a sports data tool. It will describe and discuss the play, but it does not measure speeds, distances or player positions.
- AI answers need a Google sign-in. Each account has a daily allowance.
Checking this device
- The full scan runs on your graphics card. That is 1 to 4 frames a second. It needs a browser that lets a page use the card, such as a recent Chrome or Edge, and a dedicated card or a modern built-in one.
- Where the browser cannot use the card, the same models run on the processor. That is slower, and the scan looks at fewer frames: from 1 every 4 seconds up to 2 a second.
- Phones and tablets do less on the device. So do computers with under 4 GB of memory. Scene changes are found on the device and the rest is done by deep scan and questions.
- Speed follows your graphics card. On our test laptop, with an RTX 2080 Super, the picture of a 15-minute 4K video is scanned in about 3 minutes, and 12 minutes of speech is written down in about 2 minutes.
- The accurate speech model only runs in the full scan. It is a 1.6 GB download the first time, then it is kept on your device. Where the graphics card cannot run it, or there is no room to keep it, a smaller model is used, and that one is only dependable in English, Spanish and Portuguese.
Show 3 moreShow fewer
- H.265 (HEVC) and AV1 need the right hardware. Those files only open on devices that can decode them.
- Big files may need one click after a refresh. Files up to 1.5 GB carry on by themselves. Bigger ones carry on after one click in Chrome and Edge. Other browsers ask for the file again.
- The YouTube live scan needs desktop Chrome or Edge. It also needs a graphics card the browser can use.
Methods in use
- Object detection
- A detector trained on 365 kinds of object runs on every frame the scan looks at.
- Sliced detection
- Scene keyframes are checked again as overlapping close-up tiles, to catch small and distant things.
- Object tracking
- Detections are joined from frame to frame, so each object has a start, an end and a path.
- Scene change detection
- Colour and brightness fingerprints mark where one shot ends and the next begins.
- Keyframe sampling
- Where a file's own keyframes are close enough together the scan reads those, which are its sharpest frames.
- Speech recognition
- A speech model writes the transcript, with a time on every line.
- Language detection
- The speech model works out which language is being spoken before it writes anything down.
- Voice detection
- A second model listens for a human voice, so lines written over music or noise can be marked as unclear.
- Look-again answers
- The AI model answers from frames it asks for while it works, not from a one-off description.
- Scene description
- Deep scan writes an inventory of each scene.
- On-screen text reading
- Text in the picture is read during deep scans and when answering.
- Summary and chapters
- For YouTube links, the model writes a summary, chapters and key moments.
- Search
- Everything found can be searched by word.
05Privacy
Your video stays yours.
Scanning happens on your device
The file is opened by your browser and analysed there. It is not uploaded.
Questions send frames, not the video
When you ask something or run a deep scan, only the frames needed are sent to Google's Gemini model to be read. This site keeps no copy.
Results live in your browser
Scans are saved on your device so you can come back to them. Delete them whenever you like.
Sign in only for AI answers
Google sign-in protects the AI features from abuse. There is no profile and no database behind it.
Got a video? Find out what is in it.
It works on most computers, and is fastest on one with a dedicated graphics card, the kind built for games or video work. The scanner checks your device and picks the right mode.