Analyze your Valorant gameplay with state-of-the-art vision models. Confirm suspicions in seconds with 99% confidence, all running locally on your machine.
Vanguard++ brings server-grade analysis to your desktop.
Feeds gameplay through a custom Vision Transformer, extracting 16 keyframes per suspect clip for deep analysis.
Uses librosa to automatically detect kill sounds and slice clips, so you don't have to manually edit.
Fine-tune the model on your own dataset. Label clips as "clean" or "sus" to improve accuracy for your rank.
Get a precise 0.0 to 1.0 probability score for every clip alongside frame-by-frame anomaly highlighting.
Zero cloud uploads. All inference runs locally on your GPU. Your footage never leaves your machine.
Automatically sorts clips into /cheating and /legit folders based on verification results.
Extracting frames (14/16)...
Aimbot behavior detected in frames 12-14
# Start training with your custom dataset
python main.py --train --epochs 50
Loading dataset... Found 450 clips.
Epoch 1/50 [==>........] loss: 0.452 acc: 0.720
Epoch 2/50 [====>......] loss: 0.312 acc: 0.815
Epoch 3/50 [======>....] loss: 0.188 acc: 0.902
Training complete. Model saved to /models/v2.pt
# Clone the repository
git clone https://gitlab.com/Vyapari-Dev/vanguard.git
cd vanguard
# Setup environment
python -m venv venv
.\venv\Scripts\activate
# Install dependencies
pip install -r requirements.txt
# Launch
python main.py
Required
Recommended
Automate the review of thousands of player clips to ensure competitive integrity in your brackets.
Verify your insane clips are legit before posting, or "expose" suspcious players in your lobbies.
Keep your private scrims and ten-mans clean with automated replay analysis.
Yes. Vanguard++ runs entirely locally on your machine. It analyzes video files you provide. It does not hooks into the game memory or modify game files, so it is safe and won't trigger anti-cheat bans.
Currently, it is optimized for detecting "flick" style aimbots and mechanical triggerbots that exhibit non-human reaction times or cursor movement. Wallhacks are harder to detect purely from video.
Inference is relatively lightweight, but training your own model is intensive. We recommend at least an NVIDIA GTX 1660 or better for reasonable training times.
Absolutely! The project is open source on GitLab. We welcome pull requests for new models, UI improvements, or better clip extraction logic.