Script and template as the only truth
Beats, voice lines, number cards and colours live in a script file; every layout, timing and motion constant lives in a template. A new video is a new script, not new code.
Side project · R&D · Automated 9:16 product-review videos
Review video pipeline
A local pipeline that turns a script, real product photos and a music track into a finished vertical review video, gated by automated checks.
Rendered output
- Role
- Solo: research, pipeline, QC suite
- Year
- 2026
- Platforms
- CLI
- Stack
- Python
- Pillow
- NumPy
- ffmpeg
- VieNeu-TTS
- bash
Short review videos are repetitive to make by hand, and most automated ones look cheap or invent facts. How far can a small local pipeline go if product pixels come from real photos and every on-screen claim has a source?
My part. All of it: the research report, about 7,200 lines of Python for the audio, render and QC layers, and the docs.
At a glance
25+
automated checks on every finished file
Source: QC suite
−14 LUFS
loudness target, −1 dBTP true peak
Source: audio QC
18
on-screen claims mapped to a source
Source: claims audit
What I built
Key decisions
A check is trusted only after it fails
Every QC check has a negative control that breaks the video on purpose and must be caught, and the build passes only on a fresh render.
Problem, choice and trade-off: A check is trusted only after it fails
- Problem
- Automated QC said a video passed when the report was 3.5 hours older than the render, and a layout check measured text at the origin instead of where it was drawn.
- Choice
- Every QC check got a negative control that breaks the video on purpose (text pushed out of its band, a 4.5 s shot, a subtitle declared 120 px off) and must be caught; build.sh reruns QC on the fresh render and exits 0 only when all pass.
- Why
- A green report from a check that has never been seen to fail proves nothing, and a stale report is worse than none.
- Trade-off
- Builds are slower and the negative controls need maintaining whenever a layout rule changes.
Speak one sentence, show another
One TTS call per sentence for natural phrasing and hard sync points, with the spoken form mapped to clean on-screen numbers.
Problem, choice and trade-off: Speak one sentence, show another
- Problem
- Vietnamese TTS reads numbers and units badly, and per-word timings from one long TTS call drifted from the subtitles.
- Choice
- One TTS call per sentence gives exact sentence boundaries; inside a sentence, word timings are spread by a character-weighted rule, and a mapping turns the spoken form into the displayed form ("một nghìn héc" → 1.000 Hz, "hai triệu năm chín" → 2.590.000đ).
- Why
- Sentence-level calls keep natural phrasing and give hard sync points without a forced aligner, and the viewer reads clean numbers while hearing natural speech.
- Trade-off
- Word-level karaoke timing is an estimate within each sentence, not true alignment.
Credit. Solo R&D, not a published ad. The demo reviews a Logitech mouse using Logitech's product photos for a technical demonstration only, with no affiliation. The full reference video also uses three short AI-generated clips; none appear on this page.
