Advanced AI-powered cheat detection for Counter-Strike 2 gameplay footage

This system uses cutting-edge Vision Transformer deep learning technology to analyze gameplay clips and detect potential cheating behavior from the footage. The current version is specific to CS2 and must use a user-trained model.

• Analyzes CS2 gameplay videos - Upload your raw gameplay footage or individual clips
• Focuses on killshots - Automatically extracts 2-second clips by detecting headshots from gameplay audio
• Trains locally from your footage - Select your processed clips folder and train or fine-tune a model with labels (cheating) or (not cheating)
• Analyzes processed clips - Provides detailed confidence scores from 0.0 (likely legitimate) to 1.0 (likely cheating) on clips
• Frame-by-frame analysis - See exactly which 16 frames the model analyzed for each clip
• Persistent results - All analysis results are automatically saved and can be revisited later
• Export capabilities - Export results as JSON for further analysis

1. Download this project to your computer (extract the ZIP)

2. Run wsl-install.bat as admin (not a virus trust me bro) Right click wsl-install.bat and click run as administrator. This will take a few minutes and will look frozen, but it's working in the background to install WSL.

   Once the terminal auto-closes, reboot your PC and run wsl-install.bat as admin again and follow the prompts to set up a username and password.

3. Run wsl-setup.bat NOT as admin Double click wsl-setup.bat to run it without elevated privilages. This will auto-install the conda environment and all the requirements within the new WSL you just made.

1. Download the model Go to https://huggingface.co/jinggu/jing-model/blob/main/vit_g_ps14_ak_ft_ckpt_7_clean.pth and download the .pth file

2. Place the model

   Place this .pth file in the "deepcheat/VideoMAEv2" folder found in the extracted project files. Paste any other downloaded models in "deepcheat/VideoMAEv2/output" for inferencing or fine tuning.

3. You're ready to train a model and analyze CS2 clips!

1. Start the server: Double click wsl-run.bat

2. Open your web browser and go to: http://localhost:5000

3. Follow the steps to train a model and analyze footage

1. Paste this in your terminal - Script isn't currently tested. If it doesn't work, make install.sh executable and run it. Then do the same for run.sh

echo "🚀 Starting Waldo installation..."

# Clone the repository
echo "📥 Downloading project from GitHub..."
git clone https://github.com/Mr-Homeless/waldo.git
cd waldo

# Make install.sh executable and run it
echo "🔧 Installing dependencies..."
chmod +x install.sh
./install.sh

# Download the model file
echo "🤖 Downloading AI model (1.9GB - this will take a few minutes)..."
wget --show-progress -O deepcheat/vit_g_ps14_ak_ft_ckpt_7_clean.pth \
  https://huggingface.co/jinggu/jing-model/resolve/main/vit_g_ps14_ak_ft_ckpt_7_clean.pth

# Ask user if they want to run the server
echo ""
echo "✅ Installation complete!"
echo ""
read -p "Would you like to start the server now? (y/n): " -n 1 -r
echo ""
if [[ $REPLY =~ ^[Yy]$ ]]; then
    echo "🌐 Starting server on http://localhost:5000"
    chmod +x run.sh
    ./run.sh
else
    echo "To start the server later, run: cd waldo&& ./run.sh"
fi

• Upload CS2 gameplay videos (any format: MP4, AVI, etc.)
• The system automatically detects kill moments using audio analysis
• Extracts 2-second clips around each detected kill
• Any resolution should work, but 59.94 or 60p footage is recommended
• Currently you must remove spectator kills and proximity headshot kills manually

• Label your clips as "Cheater" or "Not Cheater" with the drop down selector
• Train new models or fine-tune existing ones with your clips (currently you need at least 48 processed clips to start training)
• Real-time progress tracking with detailed json logs
• As Mr. Homeless trains more accurate models, they will be made public here for free download https://www.patreon.com/basicallyhomeless and can be placed in the deepcheat/output/ folder for use

• Select clips to analyze (from footage processed in step 1.)
• Choose a trained model for analysis
• Get detailed results including:
  • Confidence scores
  • Probability estimates using sigmoid transformation
  • Frame-by-frame analysis showing exactly what frames and ROI the model examined
  • Color-coded confidence categories

• Comprehensive dashboard with statistics and visualizations
• Individual clip analysis with video playback
• Frame viewer showing the exact 16 frames analyzed
• Export capabilities for sharing or further analysis

• Vision Transformer architecture with 1 billion parameters
• Adaptive resolution scaling - works on any video resolution (1080p, 1440p, 4K, etc.)
• Smart frame selection - analyzes frames 85-100 where killshots occur
• Center crop focus - concentrates on the crosshair area
• Clip generator functionality - Makes two second clips based on a sample sound -- change the sample sound to any "kill" sound effect for different games

• Automatic kill detection using audio pattern recognition
• Resolution-adaptive cropping maintains consistent field of view across different video qualities
• Optimal frame extraction targets the exact moments where cheating behavior is most visible

• Statistical analysis with mean, median, standard deviation
• Distribution visualization showing score patterns
• Confidence categorization with clear color coding
• Persistent storage - results saved automatically for future reference

• 0.8 - 1.0: 🔴 Very High Confidence - Likely Cheating
• 0.6 - 0.8: 🟠 High Confidence - Possible Cheating
• 0.4 - 0.6: 🟡 Medium Confidence - Uncertain
• 0.2 - 0.4: 🟢 Low Confidence - Likely Legitimate
• 0.0 - 0.2: ✅ Very Low Confidence - Likely Legitimate

The system examines:

• Crosshair movement patterns during engagements
• Reaction timing to enemy appearances
• Tracking smoothness and micro-corrections
• Pre-aim positioning before enemies are visible
• Flick accuracy and consistency patterns

• Resolution: Any (1080p, 1440p, 4K, ultrawide, etc.)
• Formats: MP4, AVI, MOV, MKV, and most common video formats
• Frame rates: 59.94, 60fps, constant frame rate recommended
• Best quality: Higher resolution and frame rate = better analysis

1. Use high-quality footage (1440p+ recommended)
2. Include multiple kills in your uploaded videos
3. Clear audio helps with automatic kill detection
4. Consistent crosshair placement in center of screen

• CPU: 12th gen intel or Ryzen 5000 or newer
• RAM: 16GB+ recommended (32GB+ for large videos)
• GPU: NVIDIA GPU with CUDA support (other GPUs may work in compatibility mode but have not been tested)
• Storage: ~10GB free space for models and temporary files

• All processing is local - your footage never leaves your computer
• No internet required for analysis (only for initial setup)
• Results stored locally in the evaluation_results folder
• Your data remains private and under your control - no cloud connection needed

• After training a new model or fine tuning a model, it may give an error code, but if it ran through the Epocs and trained, it did complete and the errors are likely not crucial to functioning.

• If running this on Windows, it will run slightly slower and sometimes look frozen/won't have output. This is normal right now.

• Auto-clipping function Works for multiple audio streams, so if its not giving you 2 second clips, double check that the game audio is in the file and in-sync with the footage.

This tool is for educational and analytical purposes.

• Results should be considered guidance, not definitive proof
• Respect privacy and competitive integrity guidelines
• Use responsibly within gaming community standards
• This is not the final, trained and tuned version of Waldo that will definitively tell you who is cheating

Current Version: Alpha 1.0 Model Version: VideoMAE v2 with CS2-specific training Last Updated: September 19th, 2025

• ☐ Create an OBS plugin that uses replay recording to record two second clips at the detected sound byte
• ☐ The wsl-setup.bat can be modified to put temp files in a better location and potentially speed up the training for Windows users
• ☐ Update pytorch without breaking everything - this should make training on newer GPUs faster
• ☐ Create a filter that deletes/removes clips from spectator view kills and proximity kills
• ☐ Have a whale lan with 500 locked down PCs to get massive amounts of clean labeled data
• ☐ Train the terminator waldo model in year 2027 on RTX 6090Tis

For questions, issues, or feedback about this beta version, please provide:

• Your system specifications
• Error messages (if any)
• Screenshots of issues
• Description of what you were trying to do

Remember: This is not finalized software. While thoroughly tested, you may encounter issues. Your feedback helps improve the system for everyone!

Happy training and analyzing! 🎯