InverityEpiron Neural Codec
Beyond H.266. Under 100 milliseconds.
Epiron beats the most efficient conventional codec ever standardized by 24% BD-rate and the strongest neural codec by 18%, measured through real entropy coding, and decodes in under a tenth of a second. Better than the quality tier on quality, nearly as fast as the speed tier on latency.
Real bitstreams, full methodology, honest caveats.
24%
BD-Rate Savings vs H.266/VVC
18%
Vs the Strongest Neural Codec
<100ms
Decode Per Image
70-100x
Faster Than Its Quality Class
The Problem
The field split in two.
For a decade, neural compression offered a choice. You could ship, or you could win. Never both.
The fast tier
Ships, but loses on quality
Lightweight models like the bmshj2018 family decode in about 60 ms, but lose to H.266/VVC on rate-distortion by 14% to 46% BD-rate. Fast enough to serve, not good enough to matter.
The quality tier
Wins, but cannot ship
Autoregressive models like mbt2018 and cheng2020 beat H.266 by 2% to 7%, then take 7 to 8 seconds to decode a single Kodak image. At 2K, roughly 41 seconds. For any pipeline serving images at volume, a non-starter.
The Result
The corner nobody else occupies.
Rate-distortion efficiency against decode latency, on equal terms. Epiron holds beyond-state-of-the-art quality at the speed tier's latency.
| Feature | Class | BD-Rate vs H.266/VVC | Decode, Kodak |
|---|---|---|---|
| Epiron | Neural, parallel decode | 24% savings | under 100ms |
| cheng2020 | Neural, autoregressive | 7% savings | ~7 - 8s |
| mbt2018 | Neural, autoregressive | 2% savings | ~7 - 8s |
| H.266/VVC (HTM 24.0) | Conventional anchor | Anchor | Encode-bound |
| bmshj2018 hyperprior | Neural, lightweight | 14% worse | ~60ms |
| bmshj2018 factorized | Neural, lightweight | 46% worse | ~60ms |
AT SCALE
It holds where it counts
At production resolution
On CLIC Professional at roughly 2K, Epiron still beats H.266/VVC by close to 1 dB at a matched operating point, while the rest of the neural field's margin shrinks. The latency gap widens instead: autoregressive decode climbs to roughly 41 seconds per image. Epiron stays in the few-hundred-millisecond range.
On the content you actually serve
Real pipelines are full of structured content: web pages, dashboards, UI, documents. Adapted to that domain, Epiron's advantage grows on both axes at once, better quality and lower bitrate than its own general-purpose configuration. That is the content class CMS and commerce pipelines serve every day.
The Pipeline
Built into Inverity's delivery layer.
Epiron is not a research demo. It is the neural reconstruction tier of a governed production pipeline.
Routed
The Neural Media Orchestrator sends each asset to the codec family that serves it best, decided per asset rather than crushed uniformly.
Verified
Every output is checked by the Perceptual Hypercube, Inverity's quality verification layer, before it ships. No asset leaves unexamined.
Delivered
Assets arrive lighter and better, with a safe classical fallback always held in reserve.
Go Deeper
The Epiron papers.
We Benchmarked Our Codec Against H.266. It Won.
Our compression model beats H.266/VVC by 24% and the strongest neural codec by 18% on Kodak, while decoding far faster. Here is the benchmark.
Inside the Epiron Benchmark: How We Beat H.266 Without the Latency Tax
A readable walkthrough of how Epiron beats H.266/VVC by 24% and the strongest neural codec by 18% on Kodak, measured through real coding, while decoding under 100ms.
What Becomes Possible When Image Quality Stops Costing You Speed
For years, better images meant slower pages. Epiron, Inverity's neural codec, removes that tradeoff, and a lot of things teams gave up on quietly become possible again.
Tired of choosing between fast and good?
If you are building anything that moves images at scale, we should talk.