Microsoft's MLVC tackles cross-platform bottleneck in neural video codecs
tanelai · reddit · 2026-07-31
While neural networks have replaced many traditional algorithms, video codecs are still dominated by hand-engineered systems like h.264/h.265. The main bottleneck is cross-platform compatibility: numerical precision differences across hardware NPUs cause entropy model decoding to fail. Current INT8 quantization or fixed-point math cannot guarantee bit-exact results.
Microsoft's proposed MLVC (Multi-platform Learned Video Codec) circumvents this by explicitly transmitting the entropy model's scale parameters via the hyperprior, removing the need for bit-exact NPU operations. It achieves 100 FPS for 360p/540p video encoding/decoding on consumer NPUs, making real-world deployment feasible.
More from Research
- Inference-free SPLADE: retrieval at BM25-like query cost without per-query inference — qdrant_engine · 2026-09-22
- Higher-resolution microscopy can hurt CNNs: downsampling 4x improves U-Net segmentation — bravo_abad · 2026-09-22
- Did OpenAI Solve the Wrong Navier-Stokes Problem? Experts Cry Loophole — joshgans · 2026-09-22
- Bridging LLM Decision Readouts into DuckDB: Zero-Token Probabilistic Classification via LuaJIT UDFs — Shoddy_Telephone9702 · 2026-09-22
- LLM agents fail to converge in double auctions, allocate less efficiently than humans — WillRinehart · 2026-09-22
- Extracting Entities and Relations from 5M Court Decisions Without an Expensive LLM Pass — SignificantZebra5883 · 2026-09-22