Skip to content
AI VIDEO PRODUCTION

Project Video Studio

Turns a script, recording or brief into an editable layered video with captions, checked audio and portable source packages.

A frame rendered by Project Video Studio: narration title over two original 3D workflow scenes
TECHNICAL REFERENCE
CURRENT SCOPE

Open-source local engine · 95 tests pass

WHO IT'S FOR

Independent build

My local video production engine for project walkthroughs and social clips. Source is public under GPL-3.0; private media, models and machine configuration are excluded.

THE ENGINEERING

A local Python engine that turns an approved script, recorded explanation or brief into an editable video project. It prepares narration or transcription, renders up to four visual layers with masks and depth-ordered callouts, composes separate portrait and landscape layouts and exports verified masters plus portable source packages. The October 21 v2 rebuild added recorded-speech timing, style checkpoints, local reference measurements, a Blender 3D adapter and safe shot reuse; the latest addition is local human matting, so text can sit behind a presenter. 95 automated tests pass; type checks, lint, Bandit, Semgrep and secret scanning pass, and GitHub CI passed all four jobs. Real 70.7-second portrait and 307.1-second landscape validation films fully decode at about -16 LUFS. It is local software with no hosted service; full style reconstruction, autonomous shot production and native editor round trips remain unfinished, and natural talking-head matting needs separate review.

PythonFFmpegBlenderONNXSQLite

BEFORE → AFTER

Real problems from this build, what I changed and the verified result.

  1. CASE 01
    Before

    Putting text behind a presenter needed foreground masks prepared by hand in other tools.

    After

    A real composition placed text behind the person and a panel in front; cancellation, deadlines and restoration were verified. Natural talking-head quality still needs review.

    How I fixed it

    Added local, hash-pinned human matting that writes synchronized foreground and mask clips and keeps the original footage in the package.

  2. CASE 02
    Before

    Reusing a rendered shot after its source or settings changed could pair new metadata with an old picture.

    After

    Only matching completed shots are reused; any changed dependency forces a re-render, and failed runs never advance the last-success pointer.

    How I fixed it

    Bound every shot to a render recipe that hashes source bytes, inputs, engine code, packages, fonts, tools and models.

  3. CASE 03
    Before

    Default portrait caption scaling covered the lower view, and a source note was hidden behind footage.

    After

    Both fixes have focused real-frame regressions; the portrait and landscape validation films fully decode with measured loudness.

    How I fixed it

    Moved captions to an explicit pixel canvas with ratio-specific safe areas and gave source notes a bounded foreground region.

Evidence reviewed: Google Doc and public repository reviewed October 25, 2026; open gates stated above