Give an idea some movement.
Move inside the input pad. Click to focus.
Watch real network predictions become motion.
01 / Your input
Arrow keys + Enter also work. Input is synthetic and stays in this browser.
02 / Camera preview
Simplified browser renderer03 / Inside the network
Each dot is an actual activation. Blue = positive, amber = negative. Connections show architecture, not feature importance.
04 / Experiment with the inputs
These are model features, not app preferences. Edits take over from pointer input until you move or trigger an action again.
Zoom network · all 20 features
3D orientation · six features
Raw outputs & implementation boundaries
Waiting for model…
Uses the exact bundled JSON weights and JavaScript ports of the dense-layer inference math. Does not run the Core ML temporal model, native event extraction, editorial priors, personalization, trajectory solver, or pixel-coverage compositor. The orientation model is a synthetic-trained bootstrap, not a guarantee of cinematic taste. No retraining or native app changes.