The detection uses the already-loaded model with minimal resources
(128x128 test frame, 5 steps), so it works even with low free RAM.
The previous 8GB RAM threshold was incorrectly skipping detection
when models were already loaded in memory.