local natural language based video blurring/anonymization tool runs on 4K at 76 fps

It's not just a text-prompt wrapper though. I benchmarked 168 combinations (7 detectors × 3 trackers × 4 skip rates × 2 resolutions) on 4K footage: Model Effective FPS on 4K What it does RF-DETR Nano Det + skip=4 76 fps Auto-detect faces/people, real-time on 4K RF-DETR Med Seg + skip=2 9 fps Pixel-precise instance segmentation masks Grounding DINO ~2 fps Text-prompted — describe what to blur Florence-2 ~2 fps Visual grounding with natural language SAM2 varies Click or draw box to select what to blur The text-prompted models (GDINO, Florence-2) are slower (~2 fps) but the flexibility is worth it — you don't need to retrain anything, just describe what you want gone. How it works locally: Grounding DINO takes your text prompt → runs zero-shot detection on each frame → ByteTrack tracks detect