Neat project, congrats on open-sourcing it! SAM3 has been instrumental in many physical AI workflows, especially in auto-labelling.<p>We’re about to launch a new real-time API for SAM3.1, that’s optimized for low latency, if folks are interested.<p>We’re seeing up to 3.5x speedup (lossless) over the native PyTorch implementation. ~40ms (ours) vs 150ms (eager-mode).<p><a href="https://vlm.run/gateway/models/facebook-sam3.1" rel="nofollow">https://vlm.run/gateway/models/facebook-sam3.1</a>