AI video pipeline
Sofa Sufferers
A short comedy series about sports fans and other everyday sufferers, made by a pipeline I built: the facts are checked before the AI draws a frame or says a word.
- 11
- episodes since 23 Oct 2026
- 3
- platforms: YouTube, Instagram, VK
- 9,928
- views on the best episode (Instagram, 9 Nov 2026)
The problem
A comedy short about a fan living through a real race has to come out while the race is still news and get every fact right: one wrong position or gap and the joke collapses into corrections in the comments. A writer, an illustrator, a voice actor and an editor are too slow and too costly for a short that comes out several times a week.
What I did
I built a pipeline one person runs in an evening: facts first, then a script checked against them, then pictures, voice and an automatic edit. The AI draws and speaks, but it never decides a number.
Engineering decisions
- 01
A fact gate before the voice
Every number and place word in the script is bound to a sourced fact. The verifier rejects the script if a value differs, a name is not backed by the facts, or a banned word, dash or exclamation slips in. Nothing goes to the voice until it reports zero findings.
- 02
One character across fifteen shots
Scenes are generated in Gemini against reference images of the hero, his flat and his props. Each frame is checked before the next one: hands, objects that rest on something, the calendar on the fridge that has to match the story.
- 03
Voice first, the edit follows it
The voiceover is recorded with a designed voice in Gemini TTS, and speech recognition (faster-whisper) gives a timing for every word. Scene cuts and subtitles follow those timings, so nothing is cut by hand.
- 04
Smooth camera, safe subtitles
The ffmpeg assembly moves the camera with sub-pixel precision, because the built-in zoom rounds to whole pixels and jitters. Subtitles stay in the lower third, clear of the platform buttons and captions, and scenes render in parallel with a worker cap that fits in memory.
Screens
Episodes, in Russian
Outcome
A running Russian-language series on YouTube, Instagram and VK: 11 episodes since 23 October 2026, the São Paulo race episode out on the evening of the race. In the four weeks to 8 November the channel’s Shorts drew 8,896 views on YouTube, and the best episode passed 9,900 views on Instagram.
Built with
- Python
- ffmpeg
- faster-whisper
- Gemini
- Gemini TTS
- Pillow
- NumPy
- OpenF1