PaperlayerView exact revision on arXiv

Source-linked AI summary

End-to-End Learned Image Coding: From Rate-Distortion Optimization to Practical Deployment

Ming Lu, Junqi Shi, Zhan Ma

arXiv:2610.22421v1eess.IVcs.MM

TL;DR

Learned image codecs have improved rate–distortion performance, while practical deployment also requires controllability, efficiency, reliability, and interoperability. This review surveys those constraints and introduces TinyLIC for controlled rate–distortion–complexity analysis. Its benchmark findings show that practical requirements carry different costs, so codec suitability extends beyond unconstrained R–D performance.

  • Problem

    Learned codecs must meet deployment constraints—including complexity, latency, rate-control flexibility, robustness, and heterogeneous-platform compatibility—that are often underrepresented in academic benchmarks.

    Source: §1 · p. 2

  • Method

    The paper reviews practical LIC capabilities and introduces TinyLIC as a common framework for rate–distortion–complexity analysis under multiple computational budgets.

    Source: §1 · p. 3, §3.4.5 · p. 19

  • Results

    The benchmark finds that practical requirements impose different costs, and the strongest unconstrained R–D codec is not necessarily the most suitable once complexity, numerical reproducibility, or communication reliability matters.

    Source: §4.3 · pp. 31–32

  • Takeaways & Limitations

    Practical codec comparison should consider rate, distortion, complexity, and reliability together rather than rely on a conventional R–D curve alone.

    Source: §4.3 · p. 32

    Rate-control schemes must balance rate accuracy, encoding complexity, and R–D performance while remaining compatible with the codec’s variable-rate mechanism.

    Source: §3.1.3 · p. 10

Abstract

from arXiv · show

End-to-end learned image coding (LIC) has achieved substantial improvements in rate--distortion (R--D) performance, yet practical deployment requires addressing a broader set of system-level constraints. This article provides a deployment-oriented review of LIC, with particular attention to the transition from high-performance neural compression models to practical codecs. We first summarize the basic end-to-end coding framework, including learned transforms, quantization, entropy modeling, and R--D optimization. We then review practical extensions that are increasingly important for deployment, including variable-rate coding and rate control, integer-precision inference and cross-platform consistency, robustness and communication resilience, lightweight and hardware-aware implementation, and standardization and interoperability. To study these issues under a common setting, we further introduce TinyLIC as an open benchmark and reproducible case study for rate--distortion--complexity analysis. Experimental results illustrate the trade-offs among compression efficiency, computational complexity, numerical consistency, and robustness. In particular, TinyLIC with approximately 10 KMACs/pixel decoding complexity substantially outperforms the H.265/HEVC image-profile anchor on Kodak, while the 50 KMACs/pixel configuration achieves performance comparable to the H.266/VVC image-profile anchor on several test sets. The review highlights practical constraints and design trade-offs that remain central to the deployment and interoperability of learned image codecs. Materials are accessible at https://njuvision.github.io/TinyLIC.

1 Introduction

Learned image coding has advanced rate–distortion performance, but practical codecs must also meet system constraints such as efficiency, rate control, robustness, and interoperability. This survey reviews those deployment needs and introduces TinyLIC as a shared benchmark for examining their trade-offs.

1 Introduction

  • Deployment requires reliable, controllable, efficient, and interoperable codecs in real communication environments, beyond improved R–D curves alone.Constraints include computational complexity, decoding latency, rate control, robustness, and compatibility with heterogeneous hardware.Source: §1 · p. 2
  • The survey centers on five coupled deployment dimensions: rate flexibility and control, numerical consistency, robustness, lightweight implementation, and standardization.These capabilities interact with compression efficiency and one another.Source: §1 · pp. 2–3
  • TinyLIC provides a lightweight, configurable reference framework for controlled rate–distortion–complexity analysis and evaluation of practical coding functions.It supports variable-rate coding, consistent decoding, adversarially robust training, progressive transmission, and network-aware loss-resilient coding.Source: §1 · p. 3
  • The review also identifies emerging directions in hardware-efficient and interoperable codecs, coding for machine intelligence, neural scene representations, and foundation-model-assisted compression..Source: §1 · p. 3

2 Core Framework: A Brief Overview

The core LIC framework replaces hand-crafted transform-coding modules with jointly optimized neural components. The review outlines its transform, quantization, entropy-modeling, and optimization components, while emphasizing that stronger R–D performance alone does not ensure practical usability.

Core framework

  • LIC replaces hand-crafted coding modules with jointly optimized neural components that map images to compact latents, discretize them, entropy-code them, and reconstruct the image.The VAE-based framework uses analysis and synthesis transforms around quantization and entropy coding.Source: §2 · p. 3, Figure 2 · p. 4

Optimization

  • The Lagrangian R–D objective combines expected coding rate with reconstruction distortion, with λ controlling the rate–quality trade-off.The expected rate is approximated by transmitted-symbol negative log-likelihood.Source: §2 · p. 4

Transform

  • Learned transforms have progressed from convolutional networks to attention, Transformer, state-space, and hierarchical designs, trading representation capacity against computational, memory, and parallelization characteristics..Source: §2 · p. 4

Quantization

  • Scalar quantization remains dominant for simplicity and entropy-coder compatibility, while more expressive vector or structured schemes add optimization or search complexity.Signal quantization discretizes latents; model quantization instead lowers network weight and activation precision for deployment.Source: §2 · pp. 4–5

Entropy modeling

  • Stronger entropy models can reduce statistical redundancy, but may add decoding steps, memory accesses, or sequential dependencies that matter for deployment.The reviewed approaches range from factorized and hyperprior distributions to autoregressive and richer context models.Source: §2 · p. 5

Performance and deployment

  • Recent LIC development has improved R–D performance, often alongside larger models and greater computation; many high-performance codecs require hundreds to over one thousand KMACs/pixel.The table is a high-level architecture trend, not a bit-exact cross-paper leaderboard, because complexity and BD-rate depend on implementation and evaluation details.Source: §2 · pp. 5–6

Practical extensions

  • Practical codec design must address bitrate control, cross-platform decoding consistency, perturbation behavior, hardware efficiency, and interoperability beyond the core R–D framework.These properties are coupled: stronger autoregressive dependencies can increase latency and error propagation, low precision can alter R–D behavior, and robustness mechanisms may trade coding efficiency for reliability.Source: §2 · p. 6

3 Practical Extensions for Deployable LIC

Practical LIC must satisfy system requirements that nominal rate–distortion performance does not fully capture. The review organizes these requirements into five related dimensions and emphasizes their interactions with the coding pipeline.

3 Practical Extensions for Deployable LIC

  • Deployable LIC must adapt to bitrate constraints, decode consistently across platforms, withstand content and transmission perturbations, fit resource budgets, and interoperate across implementations.These requirements arise at different pipeline levels and can be coupled to transforms, entropy models, and optimization objectives.Source: §3 · p. 6
  • Rate flexibility and control address bitrate or quality needs, while quantization and cross-platform consistency target low-precision execution and deterministic entropy decoding..Source: §3 · p. 6
  • Robustness, hardware-aware deployment, and standardization address perturbations and transmission errors, computational and memory costs, latency and efficiency, and common bitstream, decoder, and conformance requirements.The section considers these capabilities in relation to compression efficiency and one another, rather than as independent add-ons.Source: §3 · p. 6

3.1 Rate Flexibility and Control

The section reviews how learned image codecs provide multiple rates, progressive bitstreams, and target-rate control, while balancing R–D performance with complexity and deployment constraints. It contrasts the mechanisms and trade-offs that shape practical rate adaptation.

3.1.1 Variable-rate Coding

  • Variable-rate coding spans multiple R–D operating points in one model, unlike training a separate network for each bitrate or quality level.Learned codecs need an explicit adaptation mechanism because bitrate is coupled to latent representations, entropy statistics, and network activations.Source: §3.1 · p. 6, §3.1.1 · p. 7
  • Feature modulation and adaptive quantization retain a fixed codec topology and predictable inference complexity, while recurrent refinement and latent decomposition use sequential or discrete operating points.Rate adaptability can use all four mechanisms; modulation can provide dense or approximately continuous adaptation at nearly constant complexity.Source: §3.1.1 · pp. 7–8, Figure 3 · p. 7

3.1.2 Progressive Coding

  • Progressive coding nests operating points in one bitstream, so transmission can stop or continue refining the reconstruction without re-encoding the image.Its challenge is fine-grained truncation and stable partial reconstruction without strong dependencies or substantial R–D loss.Source: §3.1.2 · pp. 8–9

3.1.3 Rate Control

  • Rate control selects coding parameters to approach a prescribed bitrate, but the mapping from those parameters to actual rate is content-dependent and generally not analytically known.Search evaluates actual coded rates over multiple passes; predictive mapping reduces passes but risks mismatch on underrepresented content or targets.Source: §3.1.3 · pp. 9–10
  • A practical rate-control scheme must make a target operating point predictable and reachable while balancing rate accuracy, encoding complexity, R–D performance, and compatibility with the codec’s variable-rate mechanism.Repeated search improves rate matching but adds encoding passes; predictive mapping reduces computation but depends on generalization, while block-level allocation adds optimization overhead.Source: §3.1.3 · pp. 9–10

3.1.4 Discussion and Perspective

  • Practical rate adaptation requires wide, smooth rate coverage, accurate target matching, limited control overhead, and predictable complexity; spatial decisions must still meet a global bitrate budget.Local quality maps, ROI coding, and block allocation can adapt to heterogeneous content, but their decisions must satisfy the global constraint.Source: §3.1.4 · p. 10

3.2 Model Quantization and Cross-Platform Consistency

Model quantization in learned image coding serves both low-precision efficiency and deterministic codec operation across platforms. The section compares quantization strategies and explains why the required consistency level and hardware support shape deployment choices.

3.2 Model Quantization and Cross-Platform Consistency

  • Model quantization can reduce model size, memory traffic, and computation, while deterministic shared probability estimates help prevent encoder–decoder entropy desynchronization across platforms.The section distinguishes low-precision inference from the numerical agreement required for codec interoperability.Source: §3.2 · p. 11

3.2.1 Quantization-Aware Training and Post-Training Quantization

  • QAT adapts model parameters to simulated low-precision arithmetic but requires training data and another optimization stage; PTQ converts pretrained codecs without full retraining.QAT can preserve R–D performance for a prescribed format, whereas changing its precision configuration may require costly additional optimization.Source: §3.2.1 · pp. 11–12

3.2.2 Precision Granularity and Mixed-Precision Quantization

  • Mixed-precision quantization allocates bit widths according to compression sensitivity, but theoretical gains may not reduce latency when hardware handles the selected data types inefficiently.Uniform precision is straightforward and maps to common low-precision hardware, while different modules and channels can have different numerical sensitivities.Source: §3.2.2 · pp. 12–13

3.2.3 Cross-Platform Consistency

  • For studied still-image codecs, quantizing shared entropy-related modules can make entropy decoding consistent while leaving other transform components in floating point.Entropy-decoding consistency reproduces latent symbols; reconstruction consistency additionally requires reproducible decoded pixels across platforms.Source: §3.2.3 · p. 13

3.2.4 Discussion and Perspective

  • Selective quantization can target deterministic entropy decoding with limited R–D degradation, whereas broader quantization targets runtime and energy efficiency.Precision selection must also account for operator availability, memory behavior, and execution efficiency; in predictive video, small mismatches may accumulate along the prediction chain.Source: §3.2.4 · pp. 13–14

3.3 Robustness and Communication Resilience

The section treats robustness as a system-level requirement spanning adversarial inputs, repeated compression, communication errors, distribution shifts, and downstream tasks. It surveys defenses and the trade-offs between reliability and clean-condition compression performance.

3.3 Robustness and Communication Resilience

  • Learned-codec robustness covers adversarial perturbations, backdoors, repeated compression, distribution shifts, packet loss or bitstream corruption, and error propagation to downstream vision tasks.These failure modes arise at different pipeline stages and expose trade-offs between compression efficiency and reliability.Source: §3.3 · p. 14

3.3.1 Adversarial Vulnerability

  • Evasion attacks can target reconstruction distortion, coding rate, or both, while training-time backdoors use hidden triggers to cause abnormal codec behavior on triggered inputs.Perturbations may transfer from surrogate models, and attacks at the compression stage may affect downstream machine-vision systems.Source: §3.3.1 · pp. 14–15

3.3.2 Robustification Strategies

  • Robustness can be introduced through architecture, training, or inference-time defenses, each with different trade-offs in adaptation, development cost, and protection predictability.Training-based methods may reduce clean-image R–D performance; inference-time methods can be applied to existing codecs but may degrade against adaptive perturbations.Source: §3.3.2 · p. 15

3.3.3 Practical Robustness and Communication Resilience

  • Repeated compression can accumulate reconstruction drift, while packet loss can disrupt autoregressive contexts and amplify decoding errors.More independent or hierarchical representations may reduce dependencies, often at a cost in nominal R–D performance; downstream robustness must also preserve task-relevant information.Source: §3.3.3 · pp. 15–16

3.3.4 Discussion and Perspective

  • Robustness evaluation lacks common protocols across settings, making direct comparison difficult; deployment-oriented tests should report clean performance and degradation under specified impairments.Stronger protection may require additional training, redundancy, simplified entropy dependencies, or error-resilient representations, affecting bitrate, distortion, or complexity.Source: §3.3.4 · p. 16

3.4 Lightweight and Hardware-Aware Deployment

Lightweight LIC must balance compression performance with runtime and platform constraints, since high-performing models can require prohibitive computation. The section reviews complementary strategies—distillation, compression, architecture redesign, and test-time adaptation—and argues for system-level evaluation.

3.4 Lightweight and Hardware-Aware Deployment

  • Many high-performing codecs require hundreds or over 1,000 KMACs/pixel, while deployment efficiency also depends on latency, memory, parallelism, operator support, and entropy-coding overhead.Encoder and decoder complexity need not be symmetric; some applications can tolerate more encoder computation to keep decoding light.Source: §3.4 · p. 16

3.4.1 Knowledge Distillation

  • Knowledge distillation transfers coding behavior from a larger teacher to a smaller student, but retaining reconstruction alone may be insufficient when losses stem from probability modeling or latent capacity.Approaches include reconstruction- or entropy-related supervision, intermediate-feature or latent supervision, and multi-level distillation; distillation adds no inference-time modules.Source: §3.4.1 · p. 17

Quantization and Pruning

  • Quantization and structured pruning reduce precision or remove parameters and structures, but practical gains depend on support for the resulting numerical formats, tensor shapes, and operators.Structured pruning can reduce computation more directly than unstructured sparsity; quantization and pruning can also be combined.Source: Quantization and Pruning · pp. 17–18

3.4.3 Architecture-Level Design

  • FastNIC achieves BPG-level or better Kodak performance below approximately 12 KMACs/pixel encoding and 10 KMACs/pixel decoding, while AdaNIC routes computation according to image content.Jointly optimizing transform and entropy-model execution paths matters because reducing one component alone can leave another as the runtime bottleneck.Source: §3.4.3 · p. 18

3.4.4 Test-Time Adaptation

  • Test-time adaptation can improve compression by refining latents or adapting decoder parameters, but it trades additional encoding optimization or update bits for performance.This trade-off suits settings where decoding complexity and persistent model size are constrained but extra encoding effort is acceptable.Source: §3.4.4 · pp. 18–19

3.4.5 Discussion and Perspective

  • Deployment efficiency should be evaluated using actual latency, memory consumption, and target hardware alongside arithmetic complexity, because nominal MAC counts do not capture runtime behavior.Distillation, pruning, quantization, architecture redesign, and test-time adaptation offer complementary trade-offs; TinyLIC exposes multiple computational budgets under a common topology.Source: §3.4.5 · p. 19

3.5 Standardization and Interoperability

Standardization moves learned image coding toward interoperable codec systems by specifying reproducible bitstreams, decoder behavior, complexity configurations, and evaluation conditions. JPEG AI and IEEE 1857.11 illustrate different ways to address deployment needs, while stable decoding amid model evolution remains a challenge.

3.5 Standardization and Interoperability

  • Practical learned-codec standards must specify interoperable bitstreams, decoder behavior, operating profiles, complexity configurations, test conditions, and conformance requirements.Architecture, numerical precision, entropy-model execution, and hardware implementation can affect cross-platform reproducibility.Source: §3.5 · pp. 19–20, §3.5.3 · p. 21

3.5.1 JPEG AI

  • JPEG AI supports multi-complexity decoding, rate adaptation, progressive and partial decoding, compressed-domain processing, and interoperable implementation.Its decoder configurations range from approximately 8 to more than 200 KMACs/pixel, with approximately 23 KMACs/pixel demonstrated in mobile-oriented implementations.Source: §3.5.1 · p. 20

3.5.2 IEEE 1857.11

  • IEEE 1857.11 defines Base, Main, and High profiles for different complexity–performance regions and uses common evaluation conditions; reported settings show bitrate savings relative to BPG under PSNR and MS-SSIM comparisons.The Base profile emphasizes low complexity, Main balances coding efficiency and complexity, and High targets higher compression performance.Source: §3.5.2 · p. 21

3.5.3 Discussion and Perspective

  • Standardization requires stable decoder behavior despite continuing learned-codec model and functionality evolution, with future mechanisms needed to avoid fragmenting decoding compatibility.Practical adoption also depends on reproducibility and compatibility, not compression efficiency alone.Source: §3.5.3 · pp. 21–22, Figure 7 · p. 22

4 TinyLIC: A Reproducible Practical Benchmark

TinyLIC provides a shared, reproducible platform for studying deployment properties that are otherwise evaluated across different codec architectures and protocols. Its configurable design lets researchers examine rate–distortion–complexity trade-offs and practical functions against a common baseline.

4 TinyLIC: A Reproducible Practical Benchmark

  • TinyLIC is an open, reproducible benchmark with a lightweight reference architecture, configurable complexity levels, and modular support for practical coding functions.It is intended to make interactions among compression efficiency, complexity, controllability, numerical consistency, and robustness measurable within one framework.Source: §4 · p. 22
  • The framework scales its backbone across computational budgets, supports fidelity- and perceptual-realism-oriented objectives, and offers configurable deployment modules.Modules include variable-rate coding, quantization-aware consistent decoding, adversarially robust training, progressive coding, and network-aware loss resilience.Source: §4 · p. 22
  • Evaluating each practical function against a common baseline makes its bitrate, distortion, complexity, or clean-condition performance costs measurable.This supports controlled study of deployment trade-offs rather than comparisons across independently developed codecs.Source: §4 · p. 22

4.1 Design Philosophy

TinyLIC separates a compact core codec from configurable deployment tools so their costs can be studied under a shared architecture. Its lightweight transforms, controlled entropy model, complexity configurations, and optional tools support systematic rate–distortion–complexity analysis.

4.1 Design Philosophy

  • TinyLIC separates its compact core codec from configurable functions, enabling individual deployment requirements to be activated independently or together and their costs isolated.Its reusable infrastructure allows transforms, entropy models, rate control, and deployment modules to be replaced or extended within the same evaluation pipeline.Source: §4.1 · pp. 22–24

4.1.1 Efficient Architecture

  • TinyLIC’s transform uses depthwise ConvNeXt-style EDN blocks, built primarily from convolutional operations rather than computationally intensive attention modules, while its channel-autoregressive context model defaults to four steps to balance probability-modeling capacity and decoding complexity.Nano, Small, and Base configurations provide approximately 10, 20, and 50 KMACs/pixel of decoding complexity, respectively.Source: §4.1.1 · p. 24

4.1.2 Practical Tools

  • Five optional tools cover variable-rate coding, quantization-aware consistent decoding, adversarial robustness, progressive coding, and network-aware loss-resilient coding.They modify different aspects of the codec: operating point, numerical execution, training distribution, latent ordering, or entropy dependencies; progressive coding sends more informative latent elements first, while loss resilience removes channel-autoregressive dependencies.Source: §4.1.2 · pp. 24–26

4.2 Benchmark Evaluation

The benchmark evaluates TinyLIC’s rate–distortion–complexity trade-offs and deployment tools under a shared architecture, across fidelity and perceptual objectives, test sets, and model scales.

4.2.1 Experimental Setup

  • TinyLIC variants share a common architecture to evaluate deployment mechanisms under controlled conditions, reporting PSNR BD-rate against VTM Intra, LPIPS and FID BD-rate against MS-ILLM, and complexity in KMACs/pixel.Evaluation datasets are Kodak, Tecnick, CLIC Professional validation, and the JPEG AI test set.Source: §4.2 · p. 26, §4.2.1 · p. 26

4.2.2 Rate–Distortion–Perception Performance

  • On Kodak, TinyLIC-Nano’s 14.78% PSNR BD-rate relative to VTM Intra is below BPG’s 30.10%; TinyLIC-Base reaches −0.79%, −1.76%, and −4.21% on Kodak, CLIC, and Tecnick.On the JPEG AI test set, TinyLIC-Base achieves a 0.91% BD-rate; TinyLIC uses tiles there, unlike VTM Intra’s full-image processing.Source: §4.2.2 · pp. 26–27, Table 2 · p. 27
  • On Tecnick, FID BD-rate relative to MS-ILLM falls from 91.57% for TinyLIC-Nano to 35.47% for TinyLIC-Base, while TinyLIC improves over HiFiC but remains behind MS-ILLM.The same lightweight architecture supports fidelity- and realism-oriented optimization; these results are not presented as a new perceptual-compression frontier.Source: §4.2.2 · p. 27, Table 2 · p. 27

4.2.3 Effect of Practical Tools

  • Variable-rate coding and selective quantization retain performance close to their baselines, while full INT8 quantization has a larger cost: Quant-All’s Kodak BD-rate is 2.21%.Quant-Shared changes Kodak BD-rate from −0.94% for VR-Dictionary to −0.41% while maintaining cross-platform consistent decoding; variable-rate comparisons use 15 points versus four for fixed-rate evaluation.Source: §4.2.3 · p. 28
  • At a 20% training packet-loss rate, LR-GE and LR-Random reach Kodak PSNR BD-rates of 9.78% and 9.69%; removing the autoregressive entropy model trades coding efficiency for transmission robustness.Attack-resilient training has a smaller clean-image penalty than progressive and loss-resilient configurations; AR-Source and AR-Noise reach 0.75% and −0.20% on Kodak, respectively.Source: §4.2.3 · pp. 29–30

4.2.4 Visual Comparison

  • Fidelity-oriented variants preserve structural and pixel-level information, whereas perceptual-realism variants produce sharper textures at low bitrates and may synthesize high-frequency detail.Increasing complexity from Nano to Base improves both structural fidelity and visual detail.Source: Figure 10 · p. 29, §4.2.4 · p. 30

4.2.5 Model Complexity

  • Over 150 FPS at 1080p for both encoding and decoding, TinyLIC-Base reaches 35.5 FPS encoding and 49.0 FPS decoding at 4K with the optimized implementation.Nano, Small, and Base scale from approximately 10 to 50 KMACs/pixel while preserving the same codec topology; the VTM-22 speed comparison is against reference software, not production-grade VVC implementations.Source: §4.2.5 · pp. 30–31, Figure 11 · p. 31

4.3 Discussion and Perspective

The discussion frames practical learned image coding as a multi-objective design problem. TinyLIC makes trade-offs among coding efficiency, complexity, adaptability, consistency, and robustness measurable under a shared architecture.

4.3 Discussion and Perspective

  • Increasing model capacity consistently improves coding efficiency, but marginal gains decrease at higher complexities, making scalability a codec-design dimension.Nano, Small, and Base provide this scaling path without changing the fundamental codec topology.Source: §4.3 · pp. 31–32
  • Deployment features incur different costs: variable-rate coding and selective quantization have limited R–D degradation, while full integer inference, progressive coding, and loss resilience impose stronger compromises.The cost depends on where the constraint enters the coding pipeline.Source: §4.3 · p. 32
  • A codec’s best unconstrained R–D performance does not by itself determine its practical suitability, which also depends on complexity, numerical reproducibility, and communication reliability.TinyLIC is a reference for measuring application-specific trade-offs, not a single optimal solution.Source: §4.3 · p. 32

5 Concluding Remarks and Future Perspectives

The review frames deployment properties as integral to learned image coding and uses TinyLIC to evaluate practical constraints in a common setting. It then considers how coding objectives and representations may evolve, while emphasizing that future codecs must jointly address compression, deployability, and compatibility with intelligence.

5.1 Concluding Remarks

  • The review treats rate adaptability, deterministic decoding, robustness, efficiency, and interoperability as codec-design requirements, with TinyLIC providing a controlled setting to evaluate deployment constraints.The core transform–quantization–entropy–optimization framework remains the algorithmic foundation.Source: §5.1 · p. 32

5.2.1 Toward Coding for Intelligence

  • A shared compressed representation could serve human reconstruction and multiple machine tasks, but unknown downstream tasks leave open what information to preserve and how to ensure interoperability.Joint codec training with a particular detector or classifier can reduce generality and make representations sensitive to downstream-model changes.Source: §5.2.1 · pp. 32–33

5.2.2 Beyond Latent-Grid Representations

  • Implicit neural representations and scene-level primitives broaden coding beyond latent grids, but their deployment trade-offs include encoding time, new entropy models and numerical formats, and differing access and decoding behavior.INRs can support continuous spatial querying and resolution-flexible reconstruction.Source: §5.2.2 · p. 33

5.2.3 Foundation-Model-Assisted Compression

  • Foundation-model-assisted codecs could use shared generative priors for low-rate reconstruction, but plausible images may differ semantically from the source, making reliability, controllability, and semantic fidelity concerns.Evaluation may need to consider semantic fidelity and reconstruction reliability alongside bitrate and computational cost.Source: §5.2.3 · pp. 33–34

5.2.4 From Learned Image Coding to Deployable Video Coding

  • Learned video coding adds temporal error propagation and tighter latency, bandwidth, energy, and consistency constraints; rate adaptation becomes sequential, while complexity and reference dependencies limit direct transfer of image architectures.Still-image LIC is presented as a testbed for deployment techniques before they mature in learned video coding.Source: §5.2.4 · p. 34

5.3 Closing Perspective

  • The near-term challenge is reproducible, controllable, efficient, interoperable compression, while longer-term codecs may integrate machine perception, neural scene representations, and generative models.New representations and stronger priors are useful as coding technologies only if computational cost, numerical behavior, robustness, and interoperability can also be controlled.Source: §5.3 · p. 34

Declarations

The paper provides TinyLIC implementation, pretrained models, and benchmark materials, reports no competing interests, identifies funding, and lists author contributions.

Availability of data and materials

  • TinyLIC implementation, pretrained models, and benchmark materials are publicly available.The datasets used in the study are publicly available from their respective sources.Source: Availability of data and materials · p. 34

Competing interests

Funding

  • The work received partial support from National Natural Science Foundation of China Grant No. 62431011.Source: Funding · p. 35

Authors’ contributions

  • Ming Lu contributed to manuscript drafting, Junqi Shi to experimental studies and the open benchmark, and Zhan Ma to manuscript polishing and resource support.All authors reviewed and approved the manuscript.Source: Authors’ contributions · p. 35
Loading 2610.22421v1…