AI content
AI Video Pipeline
Private repository — the publicly agreed part is shownVertical recruitment videos: from storyboard to a single mp4
Problem
Video models can't hold five shot changes in one take, they mangle Cyrillic, and they keep a palette only until the first overload. Yet the output must be one file of exactly 15 seconds, 1080×1920, with legible titles.
System
A pipeline of a storyboard (5 frames, 9:16), short video-model takes and an ffmpeg assembly. The palette is pinned in prompts and negatives, titles are never generated by the model — only drawtext with a full-Cyrillic font, one line per filter (a newline renders as a tofu box). Transitions sit in the gaps between words, durations come from the actual ffprobe, not the ordered ones. When the primary video model fails with "high load", the same prompt goes to a fallback. A typo in an already finished video I fixed without regenerating: the needed letter was copied from a neighbouring word in the same frame, so font, kerning and fade matched exactly.
Role
Solo build: pipeline, prompts, assembly, post-processing.
Outcome
The build is reproducible in one command and the length lands on exactly 15.000 s. I price a scene from real billing, not the price list, and keep the most expensive model only where nothing else will do. All characters are synthetic — no cloning of real people.