How does a website
become a finished video?

The capture worker opens the public URL in a browser and reads what is visibly present. It extracts useful copy, semantic sections, calls to action, colours, font families, logo candidates, and page metadata.

It saves approved page stills and the webfonts the page actually uses where they can be retrieved safely. Hero, feature, pricing, proof, and call-to-action regions become possible source shots. Scripts, embeds, audio, video, and unsafe SVG content are not copied into the captured logo asset.

Capture preserves evidence, but it does not decide the story. The next stage chooses which of those real elements support one short argument.

The authoring stage combines the captured brand kit with a selected hook and writes a short sequence of voice lines, on-screen text, source assets, and calls to action. Each line becomes a beat with a start, an end, and a matching visual.

Voice lines are synthesised separately and measured. Their real durations set a non-negotiable lower bound for each beat. A requested rhythm may add holding time, but the timeline will not shorten a spoken line to fit an estimate.

The scene is self-contained and uses the captured local assets. Position, scale, opacity, masks, and active content are functions of timeline time rather than free-running animations.

The renderer calculates the timestamp for every frame at 30fps, sets a virtual clock to that timestamp, and calls the scene's render function. It waits for layout and paint, saves the frame, and then advances to the next timestamp.

Real capture time does not control animation time. A difficult frame can take longer to produce without causing the next frame to jump ahead. The numbered stills are encoded in order as h264 video with the yuv420p pixel format.

Vertical scenes render at 1080x1920. Horizontal scenes render at 1920x1080. Both use 30fps, so the canvas changes while the timing contract stays the same.

The measured voice files are placed at their cue positions on the final timeline. The music bed is synthesised from scratch for the render, using the selected pulse and the actual visual cut points rather than a pre-existing recording.

Voice and score are mixed together, with the score kept under the spoken track. The combined audio passes through loudness normalisation with a -14 LUFS target, then becomes a stereo AAC stream in the MP4.

A separate native sound link may be used during social publishing when that sound is eligible for the account and placement. It is not ripped into the exported master.

The finished file is probed for its expected width, height, 30fps frame rate, duration, h264 codec, yuv420p pixel format, and audio stream. Duration must stay within 100 milliseconds of the authored render length.

The scene is also inspected at sample timestamps for visible elements that extend beyond the viewport width. Separate checks decode frames from the MP4 and compare their hashes so a scene that should move cannot pass as one repeated image.

A failed technical check stops the pipeline before upload. Human review still judges the claim, crop, pacing, and legibility, but it does not have to discover a missing audio stream by accident.