Reelary
Free · runs in your browser · no upload

AI video describer: describe any video, scene by scene

Let AI describe a video for you, scene by scene. Drop in an MP4 or MOV and get a timestamped description of what is on screen in every shot, ready to copy or download as text or subtitles. Free, private, and the video never leaves your computer.

Your file is read by this browser tab and never sent anywhere. The first run downloads the AI model from Hugging Face and caches it, so later runs start faster and work offline.

How the AI video describer works

  1. Scene detection. The tool samples the video several times a second and marks a new scene wherever the picture changes sharply, the same way an editor would mark cuts.
  2. One frame per scene. It takes the frame from the middle of each scene, where the shot has settled, rather than the cut itself.
  3. Description. Florence-2, a compact vision-language model from Microsoft, writes what it sees in that frame. It runs in this browser tab through Transformers.js.
  4. Timeline. Each description is paired with the scene's start and end time, so the result works as a shot log, a subtitle file or a WebVTT descriptions track.

What people use video descriptions for

  • Shot logging. A written list of what is in a clip, so you can decide whether it is worth opening.
  • Alt text and accessibility. A first draft of descriptive text for people who cannot see the video, to be edited before publishing.
  • Video SEO and metadata. Raw material for titles, descriptions, chapters and tags that match what is actually on screen.
  • Feeding an AI assistant. Paste the timestamped descriptions into ChatGPT or Claude to summarise a video, draft a script or plan an edit, without uploading the video itself.

Choosing the level of detail

Detailed, the default, writes a short paragraph per scene: the main subject, what they are wearing or holding, the setting, colours and any text on screen. Short gives one line per scene, such as "A woman smiles while holding a pillow", and runs in roughly two thirds of the time. Use Short for a quick shot log and Detailed when the description itself is the deliverable. Without WebGPU (on the CPU) Detailed is slow, so the tool starts on Short there.

When the desktop app is the better tool

Describing one video tells you what is in it. Describing a whole footage library is what makes it searchable: once every shot has a description, "two people laughing at a kitchen table" or "drone over a red-roofed town" finds the clip without anybody having tagged it.

That is what the Reelary desktop app does in the background, on your own machine. It writes a description for every indexed shot, searches the picture itself by meaning, and lets you drop the results onto a timeline and export a cut. Read how it compares with transcript-based search.

The desktop app writes a description for every shot in your whole library and makes them searchable. Reelary is free to download for Mac and Windows; indexing and search have no file limit.

Download Reelary free How local search works

Questions

Can AI describe this video for me?

Yes. Drop the video into the tool above. It finds where each scene starts and ends, takes a frame from the middle of every scene and has a vision-language model write what it sees: the people, objects, setting, colours and action. You get a timestamped description for every scene.

Is the video uploaded to a server?

No. The video is decoded by your browser and the Florence-2 model runs inside this tab, on your graphics card through WebGPU or on the CPU. The model is downloaded once from Hugging Face and cached; your video is never sent anywhere.

Does it describe the audio or what people say?

No, it describes the picture only. To get the spoken words, run the same file through the free video to text tool, which transcribes the speech with Whisper.

How accurate are the descriptions?

They are good at the scene, the main subjects, clothing, colours, the setting and obvious actions. They do not know who a person is, cannot read small text reliably and see one frame per scene, so movement that happens between frames is not described. Treat them as a first draft to check and edit.

Can I use it for audio description or accessibility?

It is a useful starting point. Download the result as a WebVTT descriptions track or as SRT, then edit the wording for your audience. Published audio description should always be reviewed by a person.

What are the limits?

Videos up to 20 minutes and up to 40 scenes per run. The model is a one-time download of about 220 MB, or about 340 MB when your browser supports WebGPU. With WebGPU a detailed description takes one to two seconds per scene on a recent laptop; on the CPU it takes around 20 seconds, so the tool switches to short descriptions there.

Which model does it use?

Microsoft Florence-2 base, released under the MIT license, converted to ONNX and run with Transformers.js. The Reelary desktop app uses a larger local vision-language model for its scene descriptions.

More free tools

Video to text Transcribes speech in a video or audio file with OpenAI Whisper running locally in the browser, with timestamps. Exports TXT, SRT and WebVTT. No upload, no sign-up. Caption generator Generates captions for a video with Whisper in the browser, splits them into short readable lines, lets you edit them on a live preview and exports SRT, VTT or a video with the captions burned in. No upload. Extract frames from video Finds the scene changes in a video, or samples it every N seconds, and saves full-resolution JPG or PNG stills as a ZIP. Runs in the browser, no upload. Add caption to GIF Puts a caption on an animated GIF or a short clip: white caption bar, top and bottom meme text, or subtitles. Live preview and GIF export in the browser, no watermark.

Published October 10, 2026 · Models we use