Anil-matcha/AI-Youtube-Shorts-Generator

AI‑YouTube‑Shorts‑Generator – What it is

A command‑line / Python library that takes a long‑form YouTube video (or a local video file) and automatically creates short, vertical 9:16 clips ready for TikTok, Instagram Reels or YouTube Shorts. It does this by:

1. Downloading the source video.
2. Transcribing the audio with Whisper (via MuAPI or a local faster‑whisper model).
3. Running an LLM (OpenAI or Gemini, or MuAPI’s own model) over the transcript to rank moments that are likely to go viral – hooks, emotional peaks, quotable lines, etc.
4. Selecting the top‑N highlights, de‑duplicating overlapping candidates.
5. Cropping each highlight vertically (auto‑crop API or OpenCV‑based face‑tracking) and exporting MP4 files.

The result includes a short video, a viral‑score, a hook sentence and a one‑sentence reason why the clip should perform well. The whole pipeline can run entirely locally (except for the LLM call) or can be delegated to MuAPI’s hosted endpoints for a zero‑setup “API‑mode”.

---

Core Features (as described in the README)

• YouTube‑in → vertical‑short‑out – just give a URL, get one or more 9:16 MP4s.
• Two operating modes
  • --mode api (default): uses MuAPI for download, Whisper transcription, LLM ranking and auto‑crop – no local dependencies beyond Python.
  • --mode local: runs yt‑dlp, faster‑whisper, OpenCV/ffmpeg and calls OpenAI or Gemini for the highlight ranking.
• Virality‑aware highlight selection – the LLM scores clips on hooks, emotional peaks, opinion bombs, revelations, conflict, quotable lines, story peaks, practical value.
• Score, hook, and reasoning are returned for every clip.
• Chunking & deduplication for videos > 30 min so no important moment is missed.
• Custom aspect ratios (9:16, 1:1, or any user‑specified ratio).
• CLI and Python library – can be invoked from the shell or imported (generate_shorts).
• JSON output (--output‑json) containing the full transcript, candidate highlights and final clip metadata.
• Batch processing via xargs on a list of URLs.
• MIT‑licensed, self‑hostable – you keep the videos on your own hardware.

---

How it works (pipeline overview)

Step	API mode (muapi)	Local mode
Download	MuAPI /youtube-download	yt‑dlp (or direct file path)
Transcribe	MuAPI /openai‑whisper (whisper‑1)	faster‑whisper (CPU or CUDA)
Highlight ranking	MuAPI gpt‑5‑mini (virality prompt)	OpenAI gpt‑4o‑mini or Gemini gemini‑2.5‑flash via the LLM_PROVIDER env var
Vertical crop	MuAPI /autocrop (face‑tracking)	ffmpeg + OpenCV face tracking
Output	Hosted URLs	Local MP4 files

The LLM receives a transcript (timestamped JSON) and a system prompt that defines the “virality framework”. It returns a list of candidate clips with scores and explanatory text. Overlapping candidates (> 50 % overlap) are collapsed, the top‑N are kept, and each segment is cut and re‑framed vertically.

---

Quick start (self‑hosted)

# 1. Clone & set up
git clone https://github.com/SamurAIGPT/AI-Youtube-Shorts-Generator.git
cd AI-Youtube-Shorts-Generator
python3.10 -m venv venv && source venv/bin/activate
pip install -r requirements.txt          # core (API mode)
# pip install -r requirements-local.txt  # only if you plan to use --mode local

# 2. Create .env (example)
cat > .env <<EOF
MUAPI_API_KEY=your_muapi_key
# optional for local mode
OPENAI_API_KEY=your_openai_key
EOF

# 3. Generate shorts (API mode – no extra deps)
python main.py "https://www.youtube.com/watch?v=VIDEO_ID"

# 4. Or run locally (requires ffmpeg, yt‑dlp, faster‑whisper)
python main.py "https://www.youtube.com/watch?v=VIDEO_ID" --mode local --num-clips 5

---

When to use which mode

• API mode – fastest to try, no heavy dependencies, pay‑per‑clip to MuAPI. Ideal for occasional use or prototyping.
• Local mode – gives you full control, no recurring API costs, but you must install ffmpeg, yt‑dlp, faster‑whisper and have an OpenAI/Gemini key for the LLM step.

---

Pros & Cons (as inferred from the README)

Pros	Cons
Free and open‑source; only pay for the underlying LLM/Whisper APIs you choose.	Relies on external LLMs for the core highlight ranking – you need an OpenAI or Gemini key (or MuAPI credits).
No watermarks or per‑clip limits; unlimited processing on your own hardware.	Local mode can be resource‑heavy (Whisper large models, GPU needed for speed).
Fully editable virality framework – you can tweak the prompt or highlights.py.	The quality of the clips depends heavily on the LLM’s prompt; may need prompt tuning for niche content.
Supports batch processing and JSON output for automation pipelines.	Only supports YouTube URLs out‑of‑the‑box; other platforms would need custom download code.
MIT license – easy to embed in other Python projects.	No GUI; interaction is via CLI or code.

---

Who might find this useful?

• Content creators who want to repurpose long videos into Shorts without paying SaaS subscriptions.
• Agencies that need to process many videos automatically.
• Developers building a custom video‑clipping SaaS – they can use the same MuAPI endpoints or the library as a starting point.
• Researchers exploring LLM‑driven video summarisation.

---

License

MIT – you can freely use, modify and redistribute the code.