How-to films
Four short films. Narrated, captioned, and honest about what they are — every title card is drawn locally in the house typefaces, and the footage is licensed stock, credited below each film.
The pipeline, and what it refuses to do
Narration is recorded first, and the cut is built around the length of the read. The picture is stretched to fit the words — never the other way round. When a line runs longer than its footage, the last frame holds; the sentence is not trimmed to fit the shot.
Nothing here is generated by an image or video model. The title cards, section plates, lower thirds and the corner mark are drawn locally in the same typefaces and the same palette as this site, from the same globe path data as the SVG mark. A mark a model redraws is a mark that drifts.
What is not available right now. OpenAI's image, video and speech endpoints are on a billing hard limit on this account, so no generated footage was used. Narration is ElevenLabs; footage is licensed Pexels stock. If that limit is raised, the pipeline can add generated scenes — but nothing on this page is waiting on it.
Narration
ElevenLabs, one file per scene, duration measured.
Picture
Stock footage or a drawn card, cut to the measured read.
Furniture
Lower third, corner mark, flag bar composited on top.
Normalise
Same size, frame rate, pixel format and sample aspect on every segment — otherwise the concat fails outright.
Mix
One loudness pass over the finished cut, then a poster frame at 65% in.