Skip to workspace
PixelOkay
● Local processing
Tools
VIDEO TOOLS · CREATOR WORKSPACE

Scene Detector

Automatically detect visual shot changes, review and snap boundaries, and turn your footage into a thumbnail-rich shot list—all in your browser.

—Shots identified
REVIEW & EDIT

Scene timeline

▶
Every great edit starts with a clear shot listChoose your footage, then detect and refine its scenes.
00:0000:00
Change intensityNo samples yet
Frame difference Detected boundariesClick graph to seek

Shot board

Your detected scenes appear here with thumbnails and editable details.

▥No shots yet

Load a video and run detection to build a shot board.

Ready · Your files remain on this device.
04 / OPTIONAL LOCAL VISION AI

AI-assisted shot labels

After detecting shots, use a free CLIP vision model to suggest broad visual scene descriptions from sampled frames. This does not change cut detection, and suggested labels are not verified descriptions.

Video detection works without AI. Select a shot to label it.

No AI labels yet. Download and load the optional vision model using Model setup.

A similarity score is relative to the chosen candidate descriptions, not a calibrated probability. Review suggestions before applying. Model files may be downloaded from Hugging Face; your video frames are processed in this browser rather than sent to an inference API.

Built for editorial review

Manually refine missed cuts and undo or redo your boundary and annotation edits.

Creator-ready handoff

CSV, EDL, SRT and WebVTT timestamps, editable notes and a printable contact sheet help plan the edit.

Privacy by design

No account or video uploads. Optional AI downloads runtime/model assets on consent, and runs inference locally.

LOCAL AI SETUP

Set up AI on this device

First, a model needs to be loaded on your device

Scene Detector runs its optional shot labeling in your browser. Setup may download hundreds of megabytes of free model files. Subsequent sessions may reuse compatible model caches, but model memory must be loaded again. Visual scene-change detection and all original exports work without AI.

  1. 1. Check device
  2. 2. Download / cache
  3. 3. Set up vision AI
  4. 4. Label shots

Device & model resources

Measurements may be approximate
Device RAM / model memoryChecking device information…

Live RAM/GPU memory measurements are generally unavailable.

Browser website storageChecking storage…

Site-wide estimate, shared with other PixelOkay tools. This is not device RAM.

Exact-model cache has not been verified.

ONE ACTION · ALL RECOGNIZED AI MODEL DOWNLOADS

Free AI model storage across PixelOkay

Scan this website origin for WebLLM and Transformers.js model caches, plus recognizable model files in other cache buckets. This includes caches from other PixelOkay tools, not only this tool. No site data or browser databases are cleared automatically.

Run a scan to see how many model caches can be cleared together.

Unknown IndexedDB/OPFS, browser-managed caches and general application data are not deleted. Close other AI tabs first.

Having trouble downloading or loading AI?

Refresh the storage estimate and exact-model cache status first. Retry setup or choose a compatible model. If a cache is corrupted, inspect recognized model caches before any deletion.

Never clear all website data without first exporting your projects. Model-cache removal can require other PixelOkay tools to download weights again.

Checking browser capabilities…
Currently loadedNone· Local model memory is not measured by the browser

Your video frames are analyzed by a local Transformers.js vision pipeline in a dedicated browser worker. Runtime/weight assets download from a CDN and Hugging Face. No video frames are submitted to a remote inference service.

Delete recognized AI model caches?

Using Scene Detector

Build an editable shot list by comparing sampled video frames for visual changes. Correct approximate boundaries and notes, then export a handoff for editing; optional local vision AI can suggest broad labels for selected shots.

  1. Load footage and choose scan rate, sensitivity and detection mode; higher sampling inspects more frames but still does not analyze every encoded frame.
  2. Review the shot timeline, correct missed or false cuts, add titles and notes, and optionally ask the CLIP model for a label on selected still frames.
  3. Export CSV/EDL/SRT/VTT, a contact sheet or thumbnail ZIP, or a bounded set of shot clips.

Example: Scan an eight-minute product tutorial, manually add a cut that a fade obscured, rename the setup and demonstration shots, and export an EDL plus contact sheet for editorial review.

Limits and privacy: Source video is capped at 2 GiB and four hours, with at most 2,400 samples and 500 shot records. Clip ZIP export is limited to 12 shots. Cut detection is based on sampled visual differences; optional CLIP labels inspect still frames and can be wrong. AI runtime/model assets download from external hosts after setup; the video is processed locally. EDL interoperability and real-time WebM timing depend on the receiving editor and browser.

Frequently asked questions

Does AI detect the scene cuts?

The cut list comes from sampled visual differences. Optional CLIP classifies a still from a selected shot and can suggest a broad note; it does not analyze dialogue or understand the full scene.

Are exported EDL boundaries frame-accurate in every editor?

No. Detection samples are approximate, and frame rates and timecode conventions can differ. Review the exported markers against the source and the target editor’s timebase.

Related tools