How much work can disappear?
Tile reuse and atlas mapping may avoid reconstructing predictable regions. Dense changes and disocclusions still need a reliable fallback.
Investigate.
Build. Measure.
Open experiments in real-time computing. NX Warp explores how a headset can reconstruct useful pixels with less work by using information specific to rendered VR.
RESEARCH PROTOTYPE
NX Warp
A Vulkan video codec and atlas renderer developed with the custom WiVRn NX streaming stack. The work focuses on latency, reusable tiles, and the cost of reconstruction.
Inside NX Warp →THE WORKING PREMISEReuse content.
Reuse content.
Correct what changes.
Measure the tradeoffs.
THE EXPERIMENTAL PIPELINEPC → HEADSET
- 01Rendered stereo frame→
- 02Vulkan encoder→
- 03Tile transport→
- 04Vulkan decoder→
- 05Atlas & pose-aware warp→
- 06OpenXR compositor
A simplified integration diagram. Encoding, transport, reconstruction, rendering, and compositor scheduling each contribute to the result.
Open questions.
Research directions, not finished capabilities.
Where should correction go first?
Centre-first and foveated scheduling investigate where compute improves perception most. Image age and completion-time tails matter alongside throughput.
What survives a live session?
Motion artifacts, compression quality, thermal behavior, and end-to-end latency need controlled live evidence beyond a static capture.
A target is not a result.
NX Warp is pre-alpha research with visible artifacts and incomplete quality gates. Consistent 240 Hz delivery and physical motion-to-photon latency have not been demonstrated. Individual GPU timings and screenshots do not establish those outcomes.