Chapters and Scene Cuts

Chapters and Scene Cuts finds the moments where a clip changes shot, shows you each one as a thumbnail, and turns the boundaries you keep into two things you can paste or upload: a chapter list for a video description, and a WebVTT chapter track for a player.

It runs on your device. Nothing is uploaded and nothing is downloaded, because there is no model.

This is not AI, and it matters which one it is

The tool samples about one frame per second, builds a colour histogram for each sampled frame, and measures the chi-squared distance between neighbours. A distance above the sensitivity threshold is a cut. That is the whole algorithm. It is classical image statistics, it was not trained on anything, and it has no idea what is in your video.

This matters because it tells you exactly when to trust it. The detector is measuring how much the colours in the frame changed, so:

• A hard cut between two visually different shots is found reliably.
• A cut between two shots of the same room, lit the same way, may be missed, because very little colour changed.
• A camera pan across a bright window, a cut to black, or a flash can read as a cut when no edit happened there.

An earlier version of this tool shipped under an AI label. It was not AI then either.

Sensitivity

One control, and it sets the distance a frame-to-frame change has to clear before it counts as a cut.

• Fewer, longer chapters raises the threshold. Use it for an interview, a lecture, or anything shot in one place, where you want chapters at the real topic changes and not at every camera move.
• Balanced is the default and suits most edited footage.
• More, shorter chapters lowers the threshold. Use it for a fast-cut montage or a trailer where you want every shot marked.

A minimum chapter length is applied regardless of the setting, so one busy shot cannot shatter into dozens of markers. On a short clip that floor is two sampled frames; on a longer one it is five per cent of the clip.

Naming the chapters is your job

The chapters arrive numbered: Chapter 1, Chapter 2, and so on. That is deliberate. Nothing in this tool reads the picture, the audio, or a transcript, so it has nothing to name a chapter with, and a generated title would be a guess dressed as a description. Type over the numbers with what each section actually is.

The timestamps are measured. The words are yours.

The two exports

Chapter list. Plain text, one line per chapter, in the form a video description parses. Copy it or download it as a .txt.

YouTube has three published requirements for this list, and it enforces all three silently: if any one fails, your description renders as ordinary text and no chapter menu appears, with no error to tell you why. The tool checks them for you and says which one failed:

• The first timestamp must be exactly 0:00. The tool always writes 0:00 on the first line, even when the first detected boundary is a fraction of a second late.
• There must be at least three chapters.
• Every chapter must run at least ten seconds.

If a clip cannot meet those, detect again with Fewer, longer chapters to merge short shots, remove the boundaries that are making a chapter too short, or use the WebVTT file instead, which has no such rules.

WebVTT chapter track. A .vtt file with one cue per chapter. Players that support chapter tracks read it directly, and several editing and hosting tools accept it as an import.

Timestamps are floored, not rounded

A boundary measured at 12.7 seconds is written as 0:12, not 0:13. Rounding up would place the marker one second inside the next shot, so the chapter would open on a frame belonging to the previous one. Flooring keeps every marker on or just before its own cut.

Where it runs

On its own page it takes a video file and leaves it alone. In the video editor it analyses the clip you are working on and adds a Chapters tab beside Subtitles, Highlights and Privacy blur.

In both places it is read-only. It never re-encodes your video, never writes to the timeline, and adds no undo step, because it produces timestamps rather than pixels.

Limits worth knowing before you rely on it

• It samples roughly one frame per second and caps at 120 samples, so a cut is located to within about a second, and a very long clip is sampled more coarsely than a short one.
• A dissolve or a cross-fade has no single frame where the colours jump, so it is usually found late or not at all.
• It cannot distinguish a scene change from a lighting change.
• Accuracy on the thumbnail strip is the point of the review step. Look at every boundary before you publish the list.

Related

• Auto Highlights uses the same scene sampling to propose a short reel.
• Auto Subtitles transcribes the clip locally if you want chapter names drawn from what was said.