Skip to main content
This guide covers solutions to common issues you may encounter when using the AI YouTube Shorts Generator.

​
CUDA and GPU Setup

​
Issue: CUDA Libraries Not Found

If you see errors like libcudnn.so.8: cannot open shared object file, the NVIDIA CUDA libraries aren’t in your system’s library path.
1

Verify CUDA Installation

Check if CUDA libraries are installed in your virtual environment:
find venv/lib/python3.10/site-packages/nvidia -name "lib" -type d
This should return multiple paths containing CUDA libraries.
2

Set Library Path

Export the library path before running the application:
export LD_LIBRARY_PATH=$(find $(pwd)/venv/lib/python3.10/site-packages/nvidia -name "lib" -type d | paste -sd ":" -)
3

Use the run.sh Script

The provided run.sh script automatically handles this. Always run:
./run.sh "https://youtu.be/VIDEO_ID"
Instead of calling python main.py directly.
The run.sh script is the recommended way to run the application as it properly configures the CUDA environment.

​
Issue: No GPU Acceleration

If transcription is unexpectedly slow, verify GPU is being used: 
python -c "import torch; print('CUDA available:', torch.cuda.is_available())"
  • With GPU: CUDA available: True
  • CPU-only: CUDA available: False
If you have an NVIDIA GPU but see False, reinstall PyTorch with CUDA support:
pip uninstall torch
pip install torch --index-url https://download.pytorch.org/whl/cu118

​
Issue: Wrong CUDA Version

If you get compatibility errors, check your CUDA version: 
nvcc --version  # System CUDA version
python -c "import torch; print(torch.version.cuda)"  # PyTorch CUDA version
The system CUDA version and PyTorch CUDA version must be compatible. If they don’t match, reinstall PyTorch with the correct CUDA version from pytorch.org.

​
ImageMagick Policy Issues

​
Issue: No Subtitles Appearing on Video

If the video generates successfully but subtitles are missing, this is usually an ImageMagick security policy issue.
1

Check Policy File

Verify the current policy settings:
grep 'pattern="@\*"' /etc/ImageMagick-6/policy.xml
Look for a line containing pattern="@*".
2

Check Current Rights

If the output shows rights="none", ImageMagick is blocking file operations needed for subtitles.
3

Fix the Policy

Run this command to allow read/write operations:
sudo sed -i 's/rights="none" pattern="@\*"/rights="read|write" pattern="@*"/' /etc/ImageMagick-6/policy.xml
4

Verify the Fix

Confirm the change:
grep 'pattern="@\*"' /etc/ImageMagick-6/policy.xml
Should now show: rights="read|write"
This issue is specific to Linux systems. macOS and Windows ImageMagick installations typically don’t have this restriction.

​
Alternative: Manual Policy Edit

If the command above doesn’t work:
Edit /etc/ImageMagick-6/policy.xml (or /etc/ImageMagick-7/policy.xml):Find:
<policy domain="path" rights="none" pattern="@*"/>
Change to:
<policy domain="path" rights="read|write" pattern="@*"/>

​
Face Detection Failures

​
Issue: Face-Centered Crop Not Working

The application prints ✗ No face detected. Using half-width with motion tracking for screen recording even though there are faces in the video.
  1. Faces too small: Default minSize=(30, 30) may miss distant or small faces
  2. Faces not visible in first 30 frames: Detection only samples the beginning
  3. Poor lighting or unusual angles: Affects detection accuracy
  4. Low video resolution: 480p and below have less reliable face detection

​
Solution: Adjust Detection Parameters

Edit Components/FaceCrop.py at line 40:
Components/FaceCrop.py:40
# More sensitive detection
faces = face_cascade.detectMultiScale(
    gray, 
    scaleFactor=1.05,  # Was 1.1 - slower but more thorough
    minNeighbors=5,     # Was 8 - lower threshold
    minSize=(20, 20)    # Was (30, 30) - detect smaller faces
)
Lower minNeighbors values may cause false positives (detecting non-face objects as faces). Test with your typical video content.

​
Issue: Wrong Face Selected

If multiple people are in frame and the wrong person is centered: The code selects the largest face (line 43): 
best_face = max(faces, key=lambda f: f[2] * f[3])  # Largest face by area
To prioritize the face closest to center instead of the largest face, modify the selection logic to use horizontal position rather than size.

​
Transcription Issues

​
Issue: No Transcriptions Found

If you see No transcriptions found after the audio extraction step:
1

Verify Audio Extraction

Check if the audio file was created:
ls -lh audio_*.wav
If the file exists but is very small (less than 100KB), the video may not contain audio.
2

Check Audio in Source

Verify the source video has an audio track:
ffmpeg -i your_video.mp4
Look for an Audio: line in the output. If missing, the video has no audio track.
3

Check for Errors

Look for Whisper-related errors in the console output. Memory errors may indicate insufficient RAM or VRAM.
Whisper requires approximately 1-2GB of VRAM (GPU mode) or 4-8GB RAM (CPU mode) depending on the model size.

​
Issue: Transcription is Very Slow

Expected speeds:
  • GPU (CUDA): ~5-10 seconds per minute of audio
  • CPU: ~30-60 seconds per minute of audio
If transcription is much slower:
  1. Verify GPU is being used (see CUDA and GPU Setup)
  2. Check GPU memory: nvidia-smi
  3. Close other GPU-intensive applications
  4. Consider using a smaller Whisper model if VRAM is limited

​
OpenAI API Errors

​
Issue: Failed to Get Highlight from LLM

If you see ERROR: Failed to get highlight from LLM, this indicates the highlight selection failed.
  1. Invalid API Key: Check your .env file
  2. Rate Limiting: Too many requests to OpenAI API
  3. Network Issues: Connectivity problems
  4. Insufficient Credits: OpenAI account has no remaining credits
  5. Malformed Transcription: Very short videos or corrupted transcription data

​
Solution: Verify API Configuration

1

Check API Key

Verify your .env file contains a valid key:
cat .env
Should show: OPENAI_API=sk-...
2

Test API Key

Test the key directly:
curl https://api.openai.com/v1/models \
  -H "Authorization: Bearer YOUR_API_KEY"
Should return a list of available models.
3

Check Account Status

Log into platform.openai.com and verify:
  • Account has remaining credits
  • No rate limit warnings
  • API key is active
The .env file must be in the same directory as main.py. If running from a different directory, use an absolute path or copy the .env file.

​
Issue: Rate Limiting Errors

If processing multiple videos, you may hit OpenAI rate limits: 
OpenAI API error: Rate limit exceeded
Solutions:
  1. Add delays between videos in batch processing: 
    xargs -a urls.txt -I{} sh -c './run.sh {} && sleep 10'
  2. Upgrade to a higher tier OpenAI account
  3. Use a different model (e.g., gpt-3.5-turbo has higher limits)

​
Concurrent Execution Conflicts

​
Issue: File Conflicts When Running Multiple Instances

Older versions created files like audio.wav that conflicted. Current version uses session IDs.
As of the latest version, each run gets a unique session ID (8-character UUID). Temporary files are named:
  • audio_{session_id}.wav
  • temp_clip_{session_id}.mp4
  • temp_cropped_{session_id}.mp4
  • temp_subtitled_{session_id}.mp4
This allows multiple instances to run simultaneously without conflicts.

​
Verification

To verify session ID support: 
./run.sh "https://youtu.be/VIDEO_ID"
You should see: 
Session ID: a1b2c3d4
At the start of execution. Each concurrent run will have a different ID.

​
Video Quality Issues

​
Issue: Blurry or Low-Quality Output

If the final video quality is poor:
  1. Check source resolution: The output quality cannot exceed the input 
    ffmpeg -i input_video.mp4
  2. Increase bitrate in Components/Subtitles.py and Components/FaceCrop.py: 
    bitrate='5000k'  # Was '3000k'
  3. Use slower preset for better compression: 
    preset='slow'  # Was 'medium'
For 1080p source videos, use bitrate='5000k' or higher. For 720p, 3000k is usually sufficient.

​
Issue: Large Output File Sizes

If output files are too large:
  1. Lower bitrate: 
    bitrate='2000k'  # Smaller files
  2. Use faster preset (less efficient compression): 
    preset='fast'  # Faster encoding, larger files
A 2-minute 1080p vertical video typically ranges from 20-50MB depending on bitrate and content complexity.

​
Getting Additional Help

​
Collect Debugging Information

When reporting issues, include:
1

Environment Info

python --version
pip list | grep -E "torch|whisper|opencv|moviepy|langchain"
nvidia-smi  # If using GPU
2

Error Output

Copy the full console output, especially error messages and stack traces.
3

Video Details

ffmpeg -i your_video.mp4
Include resolution, duration, and codec information.

​
Alternative Solutions

Consider using the AI Clipping API which offers:
  • No installation or dependency management
  • Faster processing with optimized infrastructure
  • Better clip selection algorithms
  • Professional support

​
Community Resources

Build docs developers (and LLMs) love

Get started for freeTalk to us