Hey everyone! I’m Fredrik, the founder of Simuletic. I spent the last year building computer vision models for security, and I ran into a massive wall: Real-world data is messy, privacy-restricted, and lacks the "edge cases" that actually matter. I tried training a YOLO to detect threats and weapons but it flagged construction drills as guns. The models lacked context and good enough real world data. So I built Simuletic to fix this data issue. We use a specialized ComfyUI workflow to generate photorealistic "hard" scenarios—like drones flying in heavy snow or people holding objects that look like weapons but aren't. Today, I'm open-sourcing our "VLM-Instruct" Subset and our Weapon Detection dataset. What's inside? The weapon detection YOLO dataset contains of more than 1000+ images or highly realistic people with guns, rifles and other equipment. I have open-sourced over 100 of them as sample, to give you real valuable data to test on. Over 300 download the last month, you find it here: https://www.kaggle.com/datasets/... The VLM dataset: Grounded Reasoning: The model doesn't just say "Gun." It gives you the Bounding Box [200, 400, ...] and explains why it's a threat. Normalized Coordinates: Ready for Qwen-VL and LLaVA fine-tuning (0-1000 scale). Safety First: We include "Safe" samples to teach the model to say "No threat detected"—crucial for reducing false alarms. Who is this for? If you are building Autonomous Agents, Security Bots, or smart CCTV systems, this data is your "Ground Truth". You find it here: https://www.kaggle.com/datasets/... I’d love to hear your feedback! If anyone tests it, please give me a comment!

Simuletic

Simuletic Vision Safety: The Next Gen of Threat Detection

Simuletic Vision Safety: The Next Gen of Threat Detection