We Benchmarked Our Codec Against H.266. It Won.

For most of the last decade, the way to win an image-compression argument was to beat H.266/VVC, the most efficient conventional codec ever standardized, on rate-distortion. The catch is that the neural models which manage it tend to decode in several seconds per image, which makes them unshippable. The conventional gold standard is efficient but slow to encode; the neural challengers are efficient but slow to decode. Either way, you pay.

We built a model that doesn't make you choose. On the standard Kodak benchmark, measured through real entropy coding rather than estimated bitrate, Epiron beats H.266/VVC by 24% on BD-rate and the strongest published neural codec by roughly 18%, and it decodes in well under a tenth of a second. Better compression than the classical gold standard, better quality than the best neural model, at a latency the quality-class models can't approach.

Key Takeaways

• • On Kodak (real entropy coding, full 6-point rate curve), Epiron beats H.266/VVC by 24% BD-rate and the strongest neural incumbent by about 18%.
  • It decodes in under 100 ms, roughly 70 to 100 times faster than the autoregressive neural models in its quality class, which take about 7 seconds per image.
  • The result holds at 2K resolution and extends to a measured domain-tuning advantage on the structured, UI-and-document content that dominates real CMS and commerce pipelines.

What did we measure, and how?

We benchmarked Epiron against four widely used neural codecs and two conventional standards, on equal terms. Every number comes from real compression and decompression, the actual bitstreams written and read back, not the estimated bitrate that many papers report. The test set is Kodak (24 images) with a second pass on CLIC Professional (41 images at roughly 2K), all on a single NVIDIA A100, with decode latency timed under GPU synchronization on an idle device.

The conventional anchor is VTM 24.0, the reference implementation of H.266/VVC, run all-intra. The neural field is the standard CompressAI lineup: the bmshj2018 factorized and hyperprior models, mbt2018, and cheng2020. Fairness is enforced by comparing at matched bitrate and by computing BD-rate, the standard measure of how much bitrate one codec saves against another at equal quality. Negative BD-rate means saving rate, which means winning.

That methodology matters because the headline result lives in the gap between estimated and real coding. A model can look fast in a forward pass and then decode slowly once the real entropy coder runs, and that is exactly where the neural field's latency problem hides.

How does the field actually stack up?

The honest picture is that most neural codecs do not beat H.266, and the ones that do are too slow to ship. Against the VTM anchor on Kodak, the 2018-era bmshj models actually lose to H.266, by 46% and 14% on BD-rate. Only the heavier autoregressive models, mbt2018 and cheng2020, beat it, by about 2% and 7%. So the conventional gold standard is still ahead of most learned compression.

But those two models that beat H.266 carry a brutal latency cost. On Kodak they decode in roughly 7 to 8 seconds per image, against about 60 milliseconds for the lightweight bmshj models, a gap of more than 100 times. At 2K resolution it gets far worse: the autoregressive decode climbs to roughly 41 seconds per image, because the entropy decode runs serially and scales with pixel count. For any pipeline serving images at volume, that is a non-starter.

This is the real shape of the field. There is a fast tier that loses on quality, and a quality tier that loses on speed. Nobody occupies both corners at once. That gap is the whole opportunity.

Where does Epiron land?

In the corner nobody else occupies: better than the quality tier on quality, and nearly as fast as the speed tier on latency. On the full six-point rate curve on Kodak, measured through real coding, Epiron posts a BD-rate of 24% better than H.266/VVC and about 18% better than cheng2020, the strongest neural incumbent. At a matched operating point it sits roughly 1.3 dB above H.266, which is more than cheng2020 manages, making it the best-in-class result against the classical gold standard.

And it decodes in under 100 milliseconds. Against the only models in its quality class, the ones that take 7 seconds, that is on the order of 70 to 100 times faster. The result is not a marginal BD-rate improvement of the kind papers usually report. It is a model that holds beyond-state-of-the-art quality while decoding fast enough to actually serve.