Learned image compression has beaten JPEG on rate-distortion benchmarks for years, yet real-world adoption remains limited. That gap is not a contradiction; it is an engineering reality check.

Benchmark Wins Are Necessary but Not Sufficient

Neural codecs often deliver better PSNR and MS-SSIM at lower bitrates, especially in aggressive compression regimes. However, metric gains do not always map directly to perceived quality. Depending on content and bitrate, perceptual metrics can narrow or even invert the apparent advantage.

Why the Hyperprior Architecture Mattered

Modern neural codecs typically use an autoencoder plus entropy model. The hyperprior innovation made coding more adaptive by transmitting compact side information about latent statistics. This improved coding efficiency and established the design pattern that still dominates today.

The Three Deployment Frictions

1) Decode cost. JPEG decoding is deeply optimized and often hardware-assisted. Neural decode paths still impose higher latency and power cost, especially on edge devices.

2) Ecosystem inertia. Existing formats are supported by browsers, cameras, editors, and pipelines. Replacing that stack is expensive even when compression improves.

3) Standardization and licensing. Production teams need stable specs, predictable IP risk, and long-term interoperability. Those conditions are still maturing for many learned codecs.

Where Learned Compression Already Works

Adoption is strongest in controlled environments: server-side pipelines, domain-specific imagery, and GPU-heavy workflows. In these settings, operators can tune hardware, models, and quality targets together, making the bitrate gains economically meaningful.

The Missing Axis: Complexity

Practitioners optimize three dimensions at once: rate, distortion, and complexity. A codec that wins on quality but misses latency and power budgets will not ship. The near-term opportunity is not only better compression quality, but better quality per millisecond and per watt.