AI Clipping
Local-first pipeline that turns long videos into edited short-form clips. Completed through V13.3, with subject-aware framing planned for V14.
Status
Completed
Year
2026
Core Tech
Python, FFmpeg, faster-whisper
01 / The Problem
Automated video cropping tools often select the wrong moments, make jarring cuts, and fail to capture the pacing of human-edited clips.
02 / The Solution
A pipeline that evaluates moments, framing, and pacing to generate short vertical clips from long-form video. It is currently on version 13.3, with continuous iteration based on comparing the output quality against previous versions.
03 / How It Works
The core pipeline follows a strictly decoupled architecture:
04 / Engineering Decisions
Local-First Architecture
Built to run locally instead of relying on expensive cloud GPU APIs, utilizing faster-whisper and OpenCV for efficient local processing.
Iterative Evaluation
Instead of trusting a single LLM prompt to edit video, I iterate by comparing each generated clip to the previous version and keeping only what looks objectively better.
05 / Technology
- Python
- FFmpeg
- faster-whisper
- OpenCV
- YOLO11
07 / Learnings
- Subject-aware framing is incredibly complex due to erratic movement. Planning to integrate YOLO11 tracking in V14 to fix center-framing drift.
Ready to see it in action?