Search
Mobile menu Mobile menu
Product Management , Data science & AI , Software development Nov 19, 2026

Why Learned Image Compression Has Stalled in Production and What It Takes to Finally Change That

VECTOR Labs Team
VECTOR Labs Team
Why Learned Image Compression Has Stalled in Production and What It Takes to Finally Change That
Last updated on: Nov 19, 2026

Teams evaluating learned lossless codecs rarely reject them on compression ratio. The ratios are good, often meaningfully better than PNG or JPEG XL at equivalent quality targets. What kills the evaluation is decode speed. When a codec that produces smaller files takes seconds to open an image that a classical codec opens in milliseconds, the compression advantage becomes irrelevant to any pipeline that reads images at volume. That is the current state of learned lossless compression in 2026, and understanding why it happens structurally is the first step toward knowing when it will stop being true.

The Sequential Decoding Problem Is Architectural, Not Incidental

The latency gap between learned and classical codecs is not primarily a hardware utilisation problem. It is a design problem rooted in how learned codecs generate their probability estimates. Most learned entropy models condition each pixel or symbol on previously decoded neighbours. That dependency chain forces the decoder to proceed value by value, waiting for each prediction before it can begin the next.

This sequential constraint means that GPU parallelism, which is the obvious tool for accelerating neural inference, cannot be applied to the decoding pass in any straightforward way. The GPU sits largely idle between predictions, waiting on data that has not yet been produced. The result is that adding more compute does not proportionally reduce latency.

The arithmetic coding step compounds this. A single arithmetic decoder reading one symbol at a time creates a serial bottleneck that persists regardless of how fast the neural predictor runs. For a standard photo at 768 by 512 pixels, that is over a million sequential decode steps under conventional designs.

Why Faster Learned Codec Variants Have Not Solved the Problem

Several research directions have tried to reduce sequential steps without abandoning learned prediction entirely. Patch-based approaches group pixels and decode patches in parallel, which reduces the number of sequential steps from millions to tens or hundreds. The trade-off is that intra-patch dependencies are broken, which degrades the probability model and shrinks the compression gain.

This is the core tension that has blocked production adoption. The compression ratio of a learned codec comes from a rich, spatially aware probability model. Parallelising the decode pass requires weakening that model. Most faster variants end up in a middle ground where they are slower than classical codecs and compress less well than full learned codecs. That combination satisfies neither the speed requirements of a training data pipeline nor the size requirements of a long-term archive.

Engineering leaders evaluating these systems should treat the speed-versus-compression trade-off not as a tuning parameter but as an architectural constraint. Choosing a faster learned codec variant today typically means accepting compression ratios that are only marginally better than JPEG XL, which already runs in milliseconds on commodity hardware.

What Fixed-Step Parallel Architectures Signal About the Direction of Travel

The more promising architectural direction decouples parallelism from the probability model by redesigning the decoding order itself rather than approximating it. FMT (Fixed-step Multiscale Tree) represents this approach: it organises image data as a coarse-to-fine tree where each level contains values that are mutually independent, allowing a GPU to predict an entire level simultaneously and decode it across parallel entropy coding lanes (Ali et al., Hugging Face 2026).

The practical consequence is that decode time no longer scales with image resolution. A fixed number of sequential steps is required regardless of how many pixels the image contains, because the tree depth is bounded by the number of scales rather than the number of pixels. Ali et al. report decode speeds 9 to 61 times faster than comparable learned codecs on the same hardware, at equal or better compression ratios.

The mechanism matters here. Parallel rANS lanes carry their states between steps, so the entropy coder does not restart per symbol. This is what makes the fixed-step property hold end-to-end rather than just at the prediction stage.

What Production Image Storage Infrastructure Actually Requires

Compression ratio and decode speed are necessary but not sufficient conditions for production adoption. Infrastructure teams managing large image archives or model training pipelines also need bit-exact reproducibility, deterministic decode behaviour across hardware generations, and integration paths into existing storage and loader tooling.

Bit-exact lossless decoding is non-negotiable for training data pipelines where pixel-level consistency affects downstream model behaviour. Any codec that introduces stochastic or hardware-dependent variation in decoded values creates audit and reproducibility problems that most ML infrastructure teams are not equipped to absorb.

The write-once, read-many access pattern of image archives also means that decode throughput dominates total infrastructure cost over any realistic retention window. A codec that reduces storage footprint by 20 percent but doubles decode time will increase total compute expenditure across the archive's lifetime. Evaluations that measure compression ratio in isolation systematically underestimate this cost.

Evaluating Readiness for Your Pipeline

The criteria for evaluating a learned lossless codec in a production context differ meaningfully from a research benchmark. Throughput at batch size, not single-image latency, is the relevant metric for training data loaders. Decode behaviour under memory pressure matters more than peak-GPU performance figures. And integration cost with PyTorch or JAX data pipelines is a practical constraint that most research implementations do not address.

Fixed-step parallel architectures like FMT are at an early stage of production tooling maturity. The architectural properties are sound, and the benchmark results are credible. The gap between a compelling research result and a codec that a data engineering team can deploy without custom infrastructure work remains real and should be factored into any adoption timeline.

Teams with genuine compression urgency today should run a controlled benchmark against JPEG XL lossless on their actual hardware and access patterns before committing to a learned codec. The gap may be smaller than expected on the compression side, and larger than expected on the integration side.

FAQs

Why is decode speed more important than encode speed for most image storage use cases?

Images in training pipelines and archives are written once and read many times. Encode happens at ingest, which is a one-time cost. Decode happens every time a training job, inference pipeline, or retrieval system accesses the image. At scale, decode throughput directly determines GPU utilisation and job completion time, making it the dominant operational cost over the lifetime of an archive.

Can learned codecs be used selectively, for example only for cold storage where decode latency matters less?

Yes, and this is often the most pragmatic entry point. Cold storage archives with infrequent access can tolerate slower decode in exchange for meaningful storage cost reduction. The risk is that access patterns change: data that is cold today may become a training dataset tomorrow. Encoding into a format that creates a decode bottleneck at that point creates technical debt that is expensive to unwind at archive scale.

What does "bit-exact" lossless decoding mean in practice and why does it matter for ML pipelines?

Bit-exact lossless means that every decoded pixel is identical to the original, with no rounding, approximation, or hardware-dependent variation. For ML training pipelines, this matters because pixel-level differences between runs can affect gradient computation, data augmentation reproducibility, and experiment auditing. Any codec that introduces non-deterministic decoding behaviour makes it harder to isolate whether a model change or a data change caused a shift in training dynamics.

How should we benchmark a learned codec against JPEG XL before committing to it?

Run the benchmark on your actual hardware, with your actual image distribution, at the batch sizes your data loaders use. Single-image latency figures from research papers are measured on specific GPUs under ideal conditions and do not account for memory pressure, I/O contention, or loader parallelism. Measure end-to-end throughput in images per second at your target batch size, and measure storage cost per gigabyte of your actual dataset rather than on benchmark image sets like Kodak, which may not reflect your content characteristics.

What is the realistic adoption timeline for production-ready learned lossless codecs?

Architectural progress on fixed-step parallel decoding is credible and the performance results are meaningful. The gap is in production tooling: stable library integrations, deterministic cross-platform behaviour, and support for standard data loader interfaces. Based on the current state of research implementations, teams should plan for a 12 to 24 month window before a learned lossless codec is deployable without significant custom engineering work, and should continue evaluating JPEG XL lossless as the practical baseline in the interim.

A team that understands you
With 20+ years of experience in the world's leading consultancy companies, implementing AI and ML projects in industry-specific contexts, we are ready to hear your challenges.
Subscribe to our newsletter for insights and updates on AI and industry trends.
By clicking "Sign me up", you agree to our Privacy Policy.
By clicking the Accept button, you are giving your consent to the use of cookies when accessing this website and utilizing our services. To learn more about how cookies are used and managed, please refer to our Privacy Policy and Cookies Declaration