ai-shorts-lab 🐕📱

Turn text prompts into viral YouTube Shorts — fully automated, runs daily.

ai-shorts-lab is an end-to-end pipeline that takes a JSON storyboard of scene prompts and automatically generates, captions, formats, and uploads YouTube Shorts — no manual video editing required.

Powered by WanGP (Wan2.1 14B), FFmpeg, and the YouTube Data API v3.

---

🎬 Live Example

All videos on @GoooogleAashaan are generated entirely by this pipeline — no stock footage, no manual editing.

---

How It Works

Storyboard JSON
      │
      ▼
WanGP (Wan2.1 14B)          ← generates each scene as a video clip
      │
      ▼
FFmpeg concat                ← joins scene clips into one video
      │
      ▼
Caption overlay              ← adds text captions with FFmpeg drawtext
      │
      ▼
Portrait conversion          ← converts 832×480 landscape → 1080×1920 (9:16)
(blurred background)           with blurred background fill
      │
      ▼
YouTube upload               ← uploads as public Short via YouTube API

---

✨ Features

• Text-to-video: Write scene prompts → get a complete Short
• Auto-captions: Scene text overlaid on video
• Portrait conversion: Automatic landscape → 1080×1920 Shorts format
• YouTube API integration: Uploads directly, sets title/description/tags/privacy
• Daily automation: Run on a schedule, report results
• Low-VRAM mode: Works on 6GB GPUs with env overrides
• Batch generation: Generate and upload multiple videos in one run

---

Requirements

• OS: Windows 10/11 (WanGP requires Windows + CUDA)
• GPU: NVIDIA with 8GB+ VRAM (6GB works — see Low-VRAM Mode)
• Python: 3.10+
• FFmpeg: Installed and in PATH — download here
• WanGP: Installed separately — see WanGP Setup below
• YouTube Data API v3 credentials — see YouTube Auth

---

Installation

git clone https://github.com/lijinlar/ai-shorts-lab
cd ai-shorts-lab
python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

---

Configuration

All user-specific paths live in config.yaml. Copy the example and fill in your paths:

cp config.example.yaml config.yaml

Then edit config.yaml:

# Path to your local WanGP installation
wangp:
  dir: "C:/path/to/Wan2GP"

# Path to your YouTube OAuth2 client credentials JSON
# Download from: Google Cloud Console → APIs & Services → Credentials → OAuth 2.0 Client ID
youtube:
  oauth_credentials: "C:/path/to/youtube.oauth.json"

config.yaml is gitignored — never committed. config.example.yaml is the safe template.

---

WanGP Setup

WanGP is the video generation backend. It runs locally on your GPU using the Wan2.1 model.

1. Clone and install WanGP

git clone https://github.com/deepbeepmeep/Wan2GP
cd Wan2GP
python -m venv venv
venv\Scripts\activate
pip install -r requirements.txt

2. Download the Wan2.1 model

WanGP will prompt you to download the model on first run, or you can download manually:

• Wan2.1-T2V-14B (recommended, ~30GB) — best quality, needs 16GB VRAM or 8GB with CPU offload
• Wan2.1-T2V-1.3B (lightweight, ~5GB) — runs on 6GB VRAM, faster but lower quality

Follow the WanGP model download guide for exact steps.

3. Configure the path in ai-shorts-lab

By default, ai-shorts-lab expects WanGP at:

C:\Users\<YourUsername>\.wan2gp\Wan2GP\

To use a different path, edit WAN2GP_DIR at the top of scripts/generate_shorts_wangp.py:

WAN2GP_DIR = Path(r"C:\path\to\your\Wan2GP")

4. Verify WanGP works

Test that WanGP generates a video before running the full pipeline:

cd Wan2GP
venv\Scripts\activate
python wgp.py

This opens the WanGP UI. Generate a test clip to confirm your GPU setup is working.

---

YouTube Auth

1. Create a Google Cloud project

1. Go to Google Cloud Console
2. Create a new project
3. Enable the YouTube Data API v3
4. Go to APIs & Services → Credentials
5. Create OAuth 2.0 Client ID → Application type: Desktop app
6. Download the JSON file — save it anywhere on your machine (e.g. C:/secrets/youtube.oauth.json)
7. Set the path in your config.yaml:

   youtube:
     oauth_credentials: "C:/secrets/youtube.oauth.json"

2. Authenticate

python scripts/youtube_auth.py --channel main

This opens a browser window to authorize your YouTube account. The token is saved to out/youtube_token.json and reused for future runs.

For multiple channels:

python scripts/youtube_auth.py --channel dogs    # saves out/youtube_token_dogs.json
python scripts/youtube_auth.py --channel main    # saves out/youtube_token.json

---

Storyboard Format

Storyboards are JSON files that define what to generate. See storyboards/example.json for a full example.

{
  "default": {
    "sceneSeconds": 3,
    "fps": 24,
    "width": 480,
    "height": 832,
    "backend": "wangp",
    "upscale4k": false
  },
  "videos": [
    {
      "title": "Dog Reunites With Owner After 18 Months 😭 #shorts",
      "description": "Emotional dog reunion. #shorts #dog",
      "scenes": [
        {
          "prompt": "Extreme close-up: a golden retriever's nose twitches near a front door, amber light, static camera. Ultra-realistic, photorealistic, cinematic 4K, shallow depth of field, smooth motion, sharp focus"
        },
        {
          "prompt": "Medium wide: soldier drops duffel bag at door, retriever freezes with ears perked, tail wagging, warm hallway light, slow push-in. Ultra-realistic, photorealistic, cinematic 4K, shallow depth of field, smooth motion, sharp focus"
        }
      ]
    }
  ]
}

Prompt tips for best results

Each scene prompt should include:

1. Shot type — Extreme close-up, Medium wide, Low angle, etc.
2. Subject detail — breed, color, expression, body language
3. Action — precise movement with speed/direction
4. Environment — lighting, surfaces, time of day
5. Camera motion — static, slow push-in, tracking shot, etc.
6. Quality suffix — always end with: Ultra-realistic, photorealistic, cinematic 4K, shallow depth of field, smooth motion, sharp focus

---

Running

Generate and upload from a storyboard

python scripts/full_daily_pipeline.py --storyboard storyboards/example.json

Upload as unlisted (for testing)

python scripts/full_daily_pipeline.py --storyboard storyboards/example.json --privacy unlisted

Generate a single video directly

python scripts/generate_shorts_wangp.py storyboards/example.json out/my_video.mp4

Daily automation

Set up a Windows Task Scheduler task or any cron-compatible scheduler to run daily:

cd C:\path\to\ai-shorts-lab
.venv\Scripts\activate
python scripts/full_daily_pipeline.py --storyboard storyboards/todays_storyboard.json

---

Low-VRAM Mode

If you have a 6GB GPU or hit CUDA out-of-memory errors, set these environment variables before running:

PowerShell:

$env:WANGP_MODEL_TYPE = "t2v_1.3B"
$env:WANGP_STEPS = "10"
$env:WANGP_CFG = "4.0"

Command Prompt:

set WANGP_MODEL_TYPE=t2v_1.3B
set WANGP_STEPS=10

This switches to the 1.3B parameter model and reduces inference steps — much lower VRAM usage, slightly lower quality.

---

Project Structure

ai-shorts-lab/
├── scripts/
│   ├── full_daily_pipeline.py          # Main entry point — runs the full pipeline
│   ├── generate_shorts_wangp.py        # Core: storyboard → video via WanGP
│   ├── batch_generate_upload_series_wangp.py  # Batch multi-video generation
│   ├── add_captions.py                 # FFmpeg caption overlay
│   ├── concat_scenes.py                # FFmpeg scene concatenation
│   ├── convert_to_shorts_format.py     # Landscape → 1080×1920 portrait
│   ├── youtube_upload.py               # YouTube API upload
│   ├── youtube_auth.py                 # OAuth2 token setup
│   ├── youtube_analytics_report.py     # Performance analytics
│   ├── combine_and_upload.py           # Combine scenes + upload in one step
│   ├── wangp_generate_scene.py         # Low-level WanGP scene generator
│   ├── auto_process_wangp.py           # WanGP queue auto-processor
│   ├── daily_youtube_automation.py     # Alternate daily automation entry
│   └── archive/                        # Experimental/superseded scripts
├── storyboards/
│   └── example.json                    # Example storyboard — start here
├── out/                                # Generated videos and tokens (gitignored)
├── requirements.txt
└── README.md

---

Known Quirks

• WanGP always outputs 832×480 — the model generates landscape video regardless of the width/height params in the storyboard. The pipeline automatically converts to 1080×1920 portrait using a blurred background fill. This is expected behavior.

• No audio by default — WanGP generates silent video. Add background music separately via FFmpeg if needed.

• GPU memory spikes — each scene is generated independently to avoid OOM. Large frame counts (>81 frames) can still crash on 8GB GPUs.

• YouTube Shorts classification — YouTube automatically classifies videos as Shorts if they are ≤60 seconds and in portrait (9:16) format. Adding #shorts to the title helps.

---

License

MIT — free to use, modify, and build on.

---

Credits

• Video generation: WanGP / Wan2.1 by deepbeepmeep
• YouTube API: Google YouTube Data API v3
• Video processing: FFmpeg