MLVC aims to run cross-platform learned video codecs on consumer NPUs
If realized, MLVC could unlock real-world deployment of neural video codecs across devices, overcoming cross-hardware inconsistencies that have slowed adoption.
At a glance
- Encoding and decoding run at about 100 FPS for 360p/540p video on consumer NPUs
- MLVC transmits entropy-model scale parameters through the hyperprior to avoid cross-NPU bit-exact differences
- On Apple M3 Neural Engine, INT8 operations are simulated via FP16 rather than true INT8
- arXiv listing shows v1 on 2026-06-26 and v2 on 2026-07-29; authors include Tanel Pärnamaa and five others
The story
MLVC stands for Multi-platform Learned Video Codec for Real-World Deployment and is presented in an arXiv preprint by Tanel Pärnamaa and five other authors. The work argues for a cross-platform learned video codec designed to work across different neural processing units (NPUs) rather than rely on traditional hand‑engineered codecs.
A core challenge the paper addresses is cross-platform compatibility: small numerical differences in entropy models across NPUs can cause decoding to disagree with encoding, breaking streams. To mitigate this, MLVC transmits the entropy-model scale parameters through the hyperprior so the neural network itself doesn’t need to run bit-exactly the same across all NPUs.
The authors claim practical performance, noting both encoding and decoding run at roughly 100 frames per second for 360p/540p video on consumer NPUs, bringing learned codecs closer to deployable reality.
The paper also notes hardware-specific caveats: on the Apple M3 Neural Engine, the relevant INT8 operations are simulated using FP16 instead of a true INT8 path, meaning bit-exact results across NPUs are not guaranteed even with MLVC’s approach.
The arXiv entry lists v1 on 2026-06-26 and a revised v2 on 2026-07-29, with authoring credited to Tanel Pärnamaa and five others.