How the workbench reads a clip.

The pipeline

Every analysis follows the same six steps, in the browser tab:

1. Probe — read duration, resolution and file size; reject clips over 3 minutes or 300 MB with a clear message.
2. Sample — seek through the file at 5 fps (3 fps above 60 seconds), draw each frame to a 160×90 canvas, and compute an HSV histogram plus luma profiles.
3. Detect cuts — compare adjacent samples, then apply a persistence gate to reject fast motion and flashes.
4. Measure — for each shot, read luma, contrast, color temperature, saturation, a five-color palette, edge density, flat-block ratio, estimated motion and letterboxing.
5. Extract — grab a first and mid frame per shot at up to 1280 px, and compose a storyboard with timecodes.
6. Write — build the prompts from the measurements plus your brief, then zip the prompts, CSV, JSON, frames, storyboard and report.

Cut detection, honestly

The sampler stores a 128-bin HSV histogram for every sample and the chi-square distance to the previous sample. A naive threshold over that distance reports every fast pan as a cut. The workbench adds a second test: for a candidate at frame i, it compares the three samples before the candidate with the three after. A real cut has high similarity inside each group and low similarity between groups. A camera move has high similarity across the candidate, because the scene is continuous.

The cut time is reported to two decimals, but the practical accuracy is about ±0.1 seconds. The sampler does not decode every frame, so it does not claim frame-exact conform accuracy.

What is measured, and what is estimated

Field	Method	Confidence
Cut time	Histogram distance + persistence gate	High on hard cuts; dissolves are approximate
Luma key / contrast	5th and 95th percentile of luma	High
Color temperature	Mean red-minus-blue balance	Directional, not a white-balance reading
Palette	k-means, five clusters	Representative, not exact
Motion	1-D projection correlation between samples	Estimated: pan / tilt / handheld / static
Rhythm	Onset autocorrelation on the decoded audio	Estimated; silence yields no BPM
Letterbox	Rows and columns that are over 95% black	High when the bars are clean

The report says “estimated” where it means estimated. It never turns a measurement into a verdict about whether a shot is live, AI-generated or rendered.

From measurements to prompts

The prompt builder takes the measured camera direction, light key, contrast, color temperature, palette, duration and aspect ratio, then merges the subject, setting, style, shot size and camera override from your brief. It writes three variants: a literal remake, a same-camera / new-subject swap, and a stylized reinterpretation.

The prompts are deliberately explicit about what they do not know. If you leave the subject blank, the prompt says “your subject”, not a guessed noun. For a semantic description from the pixels, use a vision model or the open-source rea-video CLI.

The test clip

Before building the site, we ran the sampler against a 12.5-second, 1280×720 clip with seven hard cuts and deliberately fast-motion test patterns. At 5 fps it produced 63 samples in 1.65 seconds. The persistence-gated detector found 7 of 7 cuts with 0 false positives; the naive threshold version found the same seven cuts plus nine false positives. Extracting three 960×540 JPEGs took 80 ms. These are measurements from one machine and one clip, not a universal accuracy claim.

One implementation note from that test: use a Blob URL, not a local HTTP server without Range support. A no-Range server makes currentTime jump back to zero and looks like a browser bug. Production files are Blob URLs, so the issue does not affect users.

Privacy

The video is never uploaded. It is read with the File API, wrapped in a Blob URL, decoded by the browser’s video element and drawn to canvases in the tab. The ZIP is assembled in memory and downloaded. The page uses Umami for aggregate pageviews; it does not receive the video, the frames or the prompts.