Real screenshots from VisionEmoji running on iPhone
Everything runs on-device with zero cloud dependency
YOLO26m processes frames at up to 60 FPS with Neural Engine acceleration on Apple Silicon.
YOLO26m detection + per-object crop classification blends COCO and ImageNet labels.
All inference runs locally via CoreML. No data ever leaves your iPhone.
Adjust FPS, emoji scale, confidence thresholds, Kalman filter parameters, and label priority.
Smooth position tracking with configurable process and measurement noise for stable overlays.
Uses Apple's built-in emoji set rendered via NSAttributedString with NSCache optimization.