Open a video, run the analysis pass, and the backend writes a 48 kHz, 24-bit stereo WAV of impacts, footsteps and room tone anchored to the frames that caused them. Every model runs locally and nothing is uploaded.
The analysis runs twelve stages, and the order matters. Dense optical flow runs on every consecutive frame pair first, and magnitude spikes become impact frames. Only then does object detection run, and only on those frames, because running a detector on everything is how this gets slow and noisy at the same time.
Material classification runs on the objects it found, scene and room type come from keyframes, and per-object depth places each event in the room. Events are assembled per scene with type, sub-type, intensity, screen position and depth.
Progress streams to the status bar, the run is cancellable mid-stage, and temporary frames are cleaned up in a finally block rather than on the happy path.
Foley is an internal Locke Werks product. The source is private and there is no public release.
Handling of data is covered in the Locke Werks Privacy Policy. For access, or a conversation about what it would take to use this, write to archon@lockewerks.com.