Inside the Epiron Benchmark: How We Beat H.266 Without the Latency Tax

There are two ways to win an image-compression benchmark, and until now you had to pick one. You could be efficient, beating the best conventional codec on rate-distortion, but slow, because the neural models that manage it decode in seconds. Or you could be fast but less efficient, the lightweight models that ship but lose on quality. Epiron is our answer to a question the field mostly avoids: why not both?

This is the engineer's walkthrough of how we measured it, what we found, and where the honest caveats are. If you want the formal version with the full protocol, that lives in our technical report. This one is the same result, explained.

Key Takeaways

• On Kodak, measured through real entropy coding over a full rate curve, Epiron beats H.266/VVC by 24% BD-rate and the strongest neural codec by about 18%.
• It decodes in under 100 ms, roughly 70 to 100x faster than the autoregressive models in its quality class (which take ~7 seconds).
• The advantage holds at 2K and grows on structured web/UI content. We report the caveats too.

Why most neural codecs can't ship

Start with the problem Epiron solves. The neural codecs that actually beat H.266/VVC on quality, mbt2018 and cheng2020 in the standard lineup, are autoregressive: their entropy decoder runs serially, one symbol at a time, each conditioned on the last. That serialization is the latency tax. On Kodak those models decode in roughly 7 to 8 seconds per image. The lightweight models (the bmshj2018 family) decode in tens of milliseconds but lose to H.266 on rate-distortion.

So the field splits cleanly in two. There's a fast tier that isn't good enough, and a good tier that isn't fast enough. Nobody occupies both corners. And it gets worse at scale: because autoregressive decode scales with pixel count, at 2K resolution those decode times climb into the tens of seconds. For anyone serving images at volume, that's a non-starter, which is why so much of the impressive neural-compression research never makes it into a production pipeline.

How we measured it (and why it matters)

One methodological choice does most of the work here: we measured real coding, not estimated bitrate. A lot of compression papers report the estimated bitrate, the entropy of the latent under the learned prior, instead of the actual size of a written-and-read-back bitstream. Estimated bitrate skips real entropy-coder overhead and, crucially, hides decode latency, because the latency only shows up when the real arithmetic decoder runs.

Every Epiron number comes from real compression and decompression: bitstreams written to disk, read back, bitrate measured as true coded bits per pixel. We tested on Kodak (the standard 24-image reference) and CLIC Professional (41 images at ~2K, to see production-resolution behavior), all on a single A100, with decode latency timed under GPU synchronization on an idle device. The baselines are VTM 24.0 (the H.266/VVC reference), AV2's reference encoder, and the CompressAI lineup spanning the fast and slow tiers. We compare at matched bitrate using BD-rate over the full curve, the standard fair-comparison measure.

What we found

On the full rate curve on Kodak, through real coding, Epiron beats H.266/VVC by 24% on BD-rate and cheng2020, the strongest neural baseline, by about 18%. At a matched operating point it sits roughly 1.3 dB above H.266 on PSNR, more than cheng2020 manages. And it decodes in under 100 milliseconds, putting it in the lightweight tier's latency range while beating the heavyweight tier on quality.

That last part is the whole point. The improvement isn't a marginal BD-rate bump of the kind papers usually report. It's a model that holds beyond-state-of-the-art quality while decoding 70 to 100 times faster than the only models in its quality class. The ranking, best to worst on rate-distortion: Epiron, then the autoregressive neural tier, then H.266, then the lightweight tier.

It holds at 2K, and grows on real content

Kodak is small images, so we re-ran on CLIC Professional at ~2K. Epiron still beats H.266/VVC there, by close to 1 dB at a matched point, while the neural field's margin over H.266 actually shrinks at higher resolution. The latency gap widens instead of narrowing: autoregressive decode climbs into the tens of seconds, Epiron stays in the few-hundred-millisecond range.

The result that matters most for production is the domain one. Real pipelines are full of structured content, web pages, dashboards, UI, documents, whose flat regions and sharp edges look nothing like the natural photos benchmarks favor. When Epiron is adapted to that content, the advantage grows on both axes at once: better quality and lower bitrate than its own general-purpose configuration. That's the content class your CMS and commerce pipelines actually serve, which is why it's the number we care about most.

Why this matters beyond the benchmark

A 24% BD-rate win over H.266 is a strong number on its own. But the reason it matters isn't the number, it's that it comes with deployable latency. For years the choice was efficiency or speed. Epiron is evidence you don't have to choose, which changes what a production image pipeline can assume is possible.