Pipeline

shortificator is split into four stages.

1. Transcription

shortificator.transcription.transcribe() uses faster-whisper with the large-v3 model and word timestamps.

The resulting transcript is saved as:

output/{name}_transcript.json

2. Clip analysis

shortificator.analysis.llm.analyze_with_llm() sends transcript windows to Ollama and expects structured JSON candidates:

start, end, hook, reason, score

This stage is skipped entirely when you pass manual cut points with --clip — see Clip Analysis.

3. Reframing and subtitles

shortificator.rendering.short.render_short() crops the source to a 9:16 frame and burns subtitles into the video frames.

The crop behavior depends on --crop-mode. The LLM prompt depends on --content-mode.

4. Final render

FFmpeg muxes the rendered video stream with audio from the source and writes final MP4 files:

output/{name}_short_01.mp4
output/{name}_short_02.mp4

With --srt, it also writes subtitle files for the full source and for each generated Short.