metaware studios

AcousticLock 3D — vision-locked tone stabilization for acoustic guitar recording. Problem. When tracking acoustic guitar, every sway changes the mic-to-guitar distance and angle. Leaning 5 cm closer to a cardioid mic piles on bass (proximity effect); angling away dulls pick attack and air. Engineers say "sit still," players can't, so tone drifts inside a single take, and static EQ can't tell leaning from playing louder. Solution. A real-time loop that watches the player and rides the DAW's EQ for them: a camera tracks the guitar soundboard and mic in 3D, the live offset from a locked "sweet spot" streams as MIDI CCs, and Ableton's EQ Eight compensates continuously, hands-free, mid-take. How it works. (1) Phone/webcam frames feed Meta SAM 3.1, which segments "acoustic guitar" and "microphone" as text-prompted masks at a 2 FPS cadence (cached between 30 FPS video frames; background-threaded so slow cloud round-trips never stall video). (2) Meta DPT-DINOv2 (facebook/dpt-dinov2-small-nyu, local on GPU) estimates dense metric depth; masks isolate guitar/mic pixels. (3) Pinhole un-projection plus SVD/RANSAC plane fitting reduces the scene to a soundboard plane and a mic point; distance and incidence angle are measured each frame. (4) Offsets map to MIDI CC 14 (proximity low-cut), 15 (air high-shelf), 16 (gain trim) with exponential smoothing, driving EQ Eight via loopMIDI. (5) Pressing L (or footswitch) locks the current position as the target; occlusion hold freezes the last good plane through brief blockages; M/U mutes automation. Audio spectral analysis (RMS, centroid, log-spectral distance vs. reference) validates tone but never drives EQ. Control is purely geometric, so dynamics pass through untouched. Stack and compliance. Python/OpenCV/PyTorch/transformers, Behringer UMC22 plus SM57, DroidCam/phone camera, M-Vave Chocolate footswitch, Ableton Live 12. Strictly Meta-only models: SAM 3.1 plus Muse Spark (Meta API) plus DPT-DINOv2 (enforced at runtime; non-Meta checkpoints are rejected on load). Status. End-to-end on Windows/GTX 1650: live masks (around 95 percent sweet-spot match at 17.5 cm in testing), sub-second glass-to-MIDI, Ableton knobs tracking body movement.