Epiron vs JPEG AI: What a fair neural codec test shows

A benchmark built to be checked

Codec comparisons are easy to get wrong. A single flattering metric, a competitor left on a weak setting, or numbers carried over from an older run can each tilt a result without anyone noticing. We wanted the opposite: a test that holds up when someone checks it. So we put Epiron, our neural codec, against JPEG AI (ISO/IEC 6048), the first standardised learning-based image codec. Everything ran in a single session on one GPU, scored by one metric function, with JPEG AI set to its strongest configuration.

Three results are worth your time, and one honest caveat travels with them. On natural photos, Epiron uses about 41% fewer bits at equal PSNR, at parity on learned perceptual metrics. On UI and web content with domain-tuned weights, that gap widens to roughly 59%, with clear LPIPS and DISTS wins. And on encode speed, the result that matters most for a real pipeline, Epiron runs from about 27x faster than JPEG AI at its base setting to 241x faster at its high one. It does that while still winning on three of four quality metrics. The caveat: there is one metric where JPEG AI wins, and we publish it rather than bury it. A claim that cannot be wrong is not a claim. It is an advertisement, and that is the bar we set in how we benchmark.

TL;DR

• Photos (Kodak 24): Epiron uses about 41% fewer bits at equal PSNR, at parity on learned perceptual metrics (LPIPS, DISTS). Not a perceptual-superiority claim.
• UI and web content: with domain-tuned weights, about 59% fewer bits at equal PSNR, plus clear LPIPS and DISTS wins. A fixed standard cannot ship per-domain weights.
• The metric we lose on: JPEG AI wins MS-SSIM, and we report it in full rather than drop the column.
• Speed (the result that matters most): Epiron encodes about 27x faster than JPEG AI's base setting and 241x faster than its high one, and produces a file about 34% smaller than the slowest, highest-quality AVIF setting while encoding it roughly 259x faster. Smaller and faster at once.
• No human study yet. Where structural and perceptual metrics disagree this hard, a subjective test is the honest tiebreaker, and we have not run one.

The full four-metric tables, the methodology, and the integrity gates live in the technical report linked below. This page is the summary.

Two claims we will stand behind

We will defend two bit-efficiency numbers, and only in their full four-metric context. Speed is the third claim, and it stands on its own further down. The first is bit efficiency on natural photos: about 41% fewer bits than JPEG AI at equal PSNR on Kodak, measured against its maximum-effort configuration, at parity on LPIPS and DISTS. The second is the domain result, which is the one that actually separates a tunable codec from a standard.

On UI and web imagery, Epiron with domain-tuned weights posts the following against JPEG AI's strongest configuration. We show this one row here so no number travels alone. Every other comparison, including all operating points and the speed ladder, is in the report.

Negative BD-rate means Epiron spends fewer bits for equal quality. PSNR, LPIPS, and DISTS all favour Epiron here. MS-SSIM favours JPEG AI, which is the pattern discussed below. For why these metrics can point in different directions at once, see PSNR vs SSIM vs looks good to humans and why file size is the wrong metric.

Why a tunable codec beats a fixed standard

The domain result is structural, not incidental. JPEG AI is a fixed ISO specification: its weights are frozen at standardisation and cannot be retrained for your content. Epiron's weights can. Moving from generic to domain-tuned weights on UI content swings LPIPS from roughly parity to minus 22.4% and DISTS to minus 24.9% against the same maximum-effort anchor.

This is the one axis on which a standard structurally cannot compete. A committee ships one set of weights for the entire world. A tunable codec ships weights fitted to screenshots, product photography, maps, or documents. If your library is dominated by synthetic and UI content, that gap is the number that matters, and per-domain weights are the mechanism behind it. It is the same mechanism we cover in neural compression vs traditional codecs and in compressing synthetic images, charts, and UI.

The metric we lose on

MS-SSIM is the only metric where JPEG AI wins, and it disagrees with the perceptual metrics on the same images. On Kodak, MS-SSIM shows JPEG AI well ahead while LPIPS shows Epiron ahead and DISTS is near a tie. We state that plainly rather than drop the column.

We describe the divergence, we do not assign a motive. A large structural-metric lead that disappears on two independent perceptual metrics is worth noting, but confirming what it means would need a controlled subjective study we have not run. That cuts both ways: it also means we do not get to call Epiron perceptually superior on photos. When the metrics disagree this hard, a human reader study is the honest tiebreaker. This is the same trap covered in why quality 95 means nothing and our breakdown of Butteraugli and SSIMULACRA2.

Faster, not just smaller: Epiron encode speed

Bit efficiency usually costs encode time. Here it did not, and this is the result most likely to decide whether a codec is usable in production. At matched warm conditions on one A100, Epiron encodes about 27x faster than JPEG AI's base operating point and 241x faster than its high one, while still winning on three of four quality metrics against both. Against classical codecs, it is faster than cwebp and about 3.9x faster than heif-enc on the hardware each would actually deploy on.

The single strongest line combines size and speed. Epiron produces a file about 34% smaller than avifenc at its slowest, highest-quality setting, while encoding it roughly 259x faster. Smaller and faster than the maximum-effort classical option, at the same time.

One honest caveat: the two neural harnesses are not perfectly like-for-like, and JPEG AI's timings here exclude entropy coding, which favours JPEG AI. Even with that thumb on the scale for the standard, the gap is what you see above. The full ladder across four image sizes, with every caveat, is in the report.

Verifiability means naming the limits. No subjective study was run, so no perceptual-superiority claim is made in either direction. The UI result rests on one 160-image sample, not the whole population of UI imagery. The neural-versus-neural latency comparison uses harnesses that are not perfectly like-for-like. And on patents: benchmarking JPEG AI's reference software is clean, but shipping a JPEG AI implementation is a separate freedom-to-operate question this test does not touch.

That distinction, between a benchmark and a shipping decision, is the point. If you are choosing a codec for a real pipeline, honesty about the limits matters as much as any BD-rate number. It is why compression should be verifiable, and it is the standard behind our complete guide to image compression.

The full report, with all four metrics at every operating point, the methodology, the integrity gates, and the corpus disclosures, is available on request. Request the Epiron vs JPEG AI technical report.

FAQ

Is Epiron better than JPEG AI?

On bit efficiency, yes: about 41% fewer bits at equal PSNR on photos, up to 59% on tuned UI content. On encode speed, 27x to 241x faster. On perceptual metrics, at parity on photos and ahead on UI. On MS-SSIM, JPEG AI wins. Better depends on the axis, which is why we report all four plus speed.

What is JPEG AI (ISO/IEC 6048)?

JPEG AI, standardised as ISO/IEC 6048, is the JPEG committee's first learning-based image codec, using trained encoder and decoder networks rather than hand-designed transforms. It targets strong compression plus machine-vision use directly on the latent, a first for an image standard. Details are published at jpeg.org/jpegai.

Why does JPEG AI win on MS-SSIM but lose on LPIPS and DISTS?

The metrics measure different things. MS-SSIM is structural, LPIPS and DISTS are learned perceptual metrics. A large MS-SSIM lead that vanishes on both perceptual metrics is worth noting, but confirming which metric better tracks human vision here would need a subjective study we have not run.

How do you know the JPEG AI results are fair?

We ran JPEG AI at its maximum-effort configuration, the one its own published headline is quoted against, and it landed within about 1% of common test conditions. Both codecs were scored by the same metric code in the same session on one GPU. The report documents every gate.

Can I get the underlying data?

The full four-metric tables at every operating point, the rate ladder, the compute environment, and a corpus manifest are in the technical report, available on request. We quote only same-session figures, measured together, with no transferred or estimated numbers.