short-video-generator-AI: Turn a Long Video Into Shorts With AI, Free and Open Source

Finding which three minutes of a forty-minute video are worth cutting into a short is still mostly a manual scrub-and-guess job. short-video-generator-AI is a free, open-source project that automates that search: point it at a YouTube link or a local video file, and it uses a large language model to read the transcript, rank moments by how likely they are to work as a standalone clip, then cuts, subtitles, and reframes the winners to vertical without anyone touching a timeline.

The project is built by developer OStudi and lives on GitHub at short-video-generator-AI. It ships as a Python command-line tool with an optional local web interface, not a hosted app and not an editor plugin. There is no account to create and no upload queue; everything runs on the machine that runs the script.

What short-video-generator-AI actually does

The pipeline runs in four stages once a video is supplied:

Transcribe: faster-whisper produces a timestamped transcript locally, with automatic or manually forced language detection.
Classify and score: an LLM reads the transcript, first tagging the content type (podcast, interview, tutorial, vlog) and pacing, then scoring candidate segments 0 to 100 against what the project calls a "virality framework": hook strength, emotional peaks, opinions, revelations, conflict, quotable lines, story arcs, and practical value.
Deduplicate and select: overlapping high-scoring segments collapse into one, and the top N candidates (three by default, set with --n) move to rendering.
Render: each selected segment is cropped to the target aspect ratio (9:16 by default, 1:1 or a custom ratio also supported), captioned, and optionally given an AI-generated hook line at the start.

That whole sequence runs from one command:

python main.py "https://www.youtube.com/watch?v=video_id" --n 3 --ratio 9:16 --resolution 1080

A local file path works the same way, and flags exist for output resolution (360 to 1080), forcing a subtitle language, and turning off the auto-generated hook. Finished clips land in an output folder. For anyone who prefers a browser over a terminal, running server.py alongside a small local HTTP server exposes a basic web UI at localhost:8000 that wraps the same pipeline.

How the highlight detection actually works

This is the detail worth being precise about, because it is the main technical difference from other tools in the category. short-video-generator-AI does not analyze the video signal itself to find highlights. It hands the transcript text to an LLM, either OpenAI's GPT-4o-mini, Google's Gemini 2.5 Flash, or a pay-per-use provider called MuAPI, and asks that model to score narrative and emotional qualities in the language: is this a hook, an opinion, a reveal, a punchline. The candidate list, ranking, and deduplication all happen in that text-scoring step before any frame is touched.

Reframing to vertical is a separate, simpler operation: an automated crop to the chosen ratio, not a face-tracking or speaker-detection system. The README doesn't document a smart-crop or pan-between-speakers mode, so a video with a moving subject or multiple people on screen should be expected to get a static crop rather than a system that follows who's talking.

How it differs from clipify and video-autopilot-kit

Two other tools already covered here work in adjacent territory, and the differences are structural, not cosmetic.

Clipify is a Claude Code skill, not a standalone program. It runs as a /clipify slash command inside an AI coding agent, has no web UI, and its reframing step uses motion analysis on faces to decide who is speaking and pan or split-screen accordingly, purpose-built for two-person talking-head interviews. short-video-generator-AI is the opposite structurally: a standalone CLI and local web app with no dependency on any coding agent, and its crop-to-ratio reframing has no face or motion awareness at all. Where Clipify's highlight-finding reads the transcript for punchlines and pauses on-device with no external API calls, short-video-generator-AI's scoring step is delegated to a cloud LLM, which means an API key and, for two of the three supported providers, a paid bill per run.

video-autopilot-kit solves a different problem entirely. It is a production pipeline that turns scripts and raw footage into finished long-form videos, Shorts, and interview shows, built around effects, captions, and format normalization. It doesn't find highlights in existing long-form content the way short-video-generator-AI and Clipify do; it assembles new videos from source material that's already been chosen. A creator repurposing an existing back catalog of long videos is closer to short-video-generator-AI's use case than video-autopilot-kit's.

Setup and what it costs to run

Requirements are Python 3.10 or newer, on Windows, macOS, or Linux, no OS restriction the way Clipify's VideoToolbox dependency effectively limits it to Apple Silicon. Installation is a standard git clone, virtual environment, and pip install -r requirements.txt, then an .env file with credentials for whichever LLM provider is chosen. Gemini offers a free daily tier, which keeps the tool genuinely free to try; OpenAI and MuAPI are paid only. Whisper transcription runs locally regardless of which LLM handles scoring, so the only recurring cost is the highlight-detection call itself, and only if Gemini's free tier isn't used or its daily limit is exceeded. The repository is MIT licensed, so the code is fully readable and forkable.

The dependency list also marks who this fits: comfort with a terminal, a .env file, and at least one LLM API key. It is not a drag-and-drop app, but it asks for less specialized setup than Clipify's Claude Code and ffmpeg requirements, and it works directly from a YouTube URL, which Clipify does not.

Who this is for

The project description frames it plainly as a tool for turning YouTube videos into viral short clips, and the virality-scoring language in the README backs that up: it's built for creators repurposing long-form YouTube content, podcasts, interviews, and vlogs into TikTok, Reels, and Shorts, rather than for editors who need frame-accurate control over which moment gets cut. Anyone who needs speaker-aware reframing for a two-person interview is better served by Clipify; anyone assembling finished video from scripts and B-roll is better served by video-autopilot-kit.

Current traction

As of today, the repository shows 1,054 stars and 10 forks on GitHub, created July 1, 2026 and last pushed August 8. It's licensed under MIT. Star growth has been fast for a two-month-old project, roughly doubling from the 452 stars first logged in scans just weeks ago, and the open-issue count sits at zero.

Source: github.com/OStudi/short-video-generator-AI