v1.0.1
A video can only be addressed along one axis.
Seek to a timestamp and you get whatever was recorded there. You cannot ask for the same subject turned slightly further, or the same moment at a different level of detail, because the file has no notion of subject or detail — only frames. MS3E compiles media into a structure that does, then reconstructs from a state vector.
Two failure modes this sits between
Linear media is perfectly faithful and completely rigid. Every pixel is real, and the only question you can ask is when.
Direct pixel generation answers any question and guarantees nothing. Identity drifts between frames, and there is no structure to hold it steady because the model is producing appearance, not a subject that has an appearance.
MS3E takes the position that the useful object in between is an explicit structure: an identity skeleton that does not move, deformation layered on top of it, regions that own their own detail, and a spectral decomposition that separates what must be preserved from what can be regenerated.
What the state vector addresses
These are the axes the shipped scene actually declares. Changing one marks a set of nodes dirty; only those get rebuilt.
- headYaw
- headPitch
- eyeYaw
- eyePitch
- expression
- accentHue
- quality
- frequencyGain
What this is and is not
- A research MVP, not a production video codec. It is the reference implementation of the whitepaper's model.
- Residual error reduction of 99.210% (Spectral Portrait) and 97.421% (Amber Sentinel) is measured on these two asset packages, not a general claim.
- Two characters, both derived from imported baseline media. Nothing here is trained; the structure is extracted and hand-tuned.
- Browser acceptance was run in Chrome 144 on desktop. Mobile GPU paths are untested.
- The v0.1 JSON Schema in the repository describes the v0.1 scene format, which v1.0 no longer uses. There is currently no published schema for the v1.0 format.