pith. machine review for the scientific record. sign in

arxiv: 2605.05148 · v1 · submitted 2026-05-06 · 💻 cs.CV · cs.AI· cs.LG

Recognition: unknown

What Matters in Practical Learned Image Compression

Kedar Tatwawadi , Parisa Rahimzadeh , Zhanghao Sun , Zhiqi Chen , Ziyun Yang , Sanjay Nair , Divija Hasteer , Oren Rippel
Authors on Pith no claims yet

Pith reviewed 2026-05-08 16:57 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.LG
keywords learned image compressionperceptual qualityneural architecture searchon-device inferencesubjective evaluationbitrate savingsruntime optimization
0
0 comments X

The pith

By studying modeling choices and running performance-aware architecture search, a learned image codec reaches 2.3-3x bitrate savings over traditional standards while encoding 12-megapixel images in 230 ms on a phone.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper examines which design decisions control the tradeoff between perceptual quality and runtime in learned image compression. It performs ablations on multiple novel and existing techniques then searches millions of backbone configurations to find models that hit target device speeds while scoring highest on perceptual metrics. The resulting codec is shown through subjective tests to require far fewer bits than AV1, VVC, and prior learned methods for the same visual quality. If the approach holds, learned codecs become deployable on consumer hardware without sacrificing either speed or appearance.

Core claim

Joint optimization of perceptual metrics and runtime constraints, achieved by exhaustive ablations followed by performance-aware neural architecture search over backbone variants, produces a practical codec that delivers 2.3-3 times lower bitrates than AV1, AV2, VVC, ECM, and JPEG-AI at equal subjective quality and 20-40 percent savings versus the strongest prior learned codecs, with 230 ms encoding and 150 ms decoding for 12 MP images on an iPhone 17 Pro Max.

What carries the argument

Performance-aware neural architecture search over millions of backbone configurations that simultaneously maximizes perceptual metric scores and meets explicit on-device runtime targets.

If this is right

  • Learned codecs can now meet the speed and quality requirements for on-device capture and transmission of high-resolution photographs.
  • Bandwidth usage in image-centric mobile applications can be cut by more than half without visible quality loss.
  • Architecture search guided by both perceptual scores and latency becomes a repeatable method for producing practical learned codecs.
  • Traditional hand-crafted codecs face direct competition from optimized learned systems on consumer hardware.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same search-and-ablation workflow could be applied to video to close similar gaps between learned and conventional codecs.
  • Hardware-specific runtime targets in the search loop make it likely that the resulting models generalize poorly to very different accelerators without retuning.
  • If perceptual metrics continue to improve, future iterations of this codec family could widen the bitrate advantage still further.

Load-bearing premise

Subjective user studies and the selected perceptual metrics accurately reflect human visual preference across varied real-world image content and that the architecture search did not overfit to the test sets.

What would settle it

An independent subjective test on a fresh, diverse image collection that finds no bitrate reduction relative to VVC or AV1 at matched quality, or that measures encode times above 230 ms on the same phone hardware, would falsify the performance claims.

Figures

Figures reproduced from arXiv: 2605.05148 by Divija Hasteer, Kedar Tatwawadi, Oren Rippel, Parisa Rahimzadeh, Sanjay Nair, Zhanghao Sun, Zhiqi Chen, Ziyun Yang.

Figure 1
Figure 1. Figure 1: Comparisons of state-of-the-art traditional and learned view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparisons of reconstruction quality for equal filesize/BPP (bits-per-pixel). view at source ↗
Figure 3
Figure 3. Figure 3: The overall model architecture. Individual components described in Sections view at source ↗
Figure 4
Figure 4. Figure 4: Detailed architecture of the outer decoder (see Appendix view at source ↗
Figure 5
Figure 5. Figure 5: We perform neural architecture search for the outer decoder, progressively filtering the search space down from 1.4M model view at source ↗
Figure 6
Figure 6. Figure 6: Rate-distortion curves of top traditional and learned codecs, based on Elo scores (higher is better) from a large-scale subjective view at source ↗
Figure 7
Figure 7. Figure 7: Ablations on artifact-specific mitigation strategies. All images are encoded at a BPP 0.20. view at source ↗
Figure 8
Figure 8. Figure 8: presents curves for additional metrics. Although the perceptual codecs PICO, HiFiC, C3-WD and CDC sub￾stantially outperform the non-perceptual codecs based on human ratings and the perceptually-oriented objective met￾rics, they do not perform well on PSNR. Conversely, the best-performers on PSNR — DCVC-RT, TCM, ECM, and VVC — perform poorly on perceptual metrics, and require 2-3 times the bitrate to achiev… view at source ↗
Figure 9
Figure 9. Figure 9: Subjective and objective curves for the Kodak dataset. view at source ↗
Figure 10
Figure 10. Figure 10: Subjective and objective curves for the DIV2K dataset. view at source ↗
Figure 11
Figure 11. Figure 11: Architectures of the outer encoder and scale decoder. See the main paper body for details. view at source ↗
Figure 12
Figure 12. Figure 12: Screenshot of the subjective study interface as seen by the human raters view at source ↗
read the original abstract

One of the major differentiators unlocked by learned codecs relative to their hard-coded traditional counterparts is their ability to be optimized directly to appeal to the human visual system. Despite this potential, a perceptual yet practical image codec is yet to be proposed. In this work, we aim to close this gap. We conduct a comprehensive study of the key modeling choices that govern the design of a practical learned image codec, jointly optimized for perceptual quality and runtime -- including within the ablations several novel techniques. We then perform performance-aware neural architecture search over millions of backbone configurations to identify models that achieve the target on-device runtime while maximizing compression performance as captured by perceptual metrics. We combine the various optimizations to construct a new codec that achieves a significantly improved tradeoff between speed and perceptual quality. Based on rigorous subjective user studies, it provides 2.3-3x bitrate savings against AV1, AV2, VVC, ECM and JPEG-AI, and 20-40% bitrate savings against the best learned codec alternatives. At the same time, on an iPhone 17 Pro Max, it encodes 12MP images as fast as 230ms, and decodes them in 150ms -- faster than most top ML-based codecs run on a V100 GPU.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The manuscript presents a comprehensive empirical study of modeling choices for practical learned image codecs jointly optimized for perceptual quality and runtime. It includes ablations of several novel techniques, followed by performance-aware neural architecture search (NAS) over millions of backbone configurations to identify models that meet on-device runtime targets while maximizing perceptual metrics. The resulting codec is claimed to deliver 2.3-3x bitrate savings versus AV1, AV2, VVC, ECM and JPEG-AI, and 20-40% savings versus prior learned codecs, according to subjective user studies, while achieving 230 ms encode / 150 ms decode for 12 MP images on an iPhone 17 Pro Max.

Significance. If the reported gains are reproducible and generalizable, the work would be significant for the learned compression community by demonstrating a deployable perceptual codec that substantially outperforms both traditional standards and existing learned methods in bitrate efficiency while satisfying strict mobile runtime constraints. The scale of the ablations and NAS could provide reusable insights for future architecture design under perceptual and efficiency objectives.

major comments (3)
  1. [Architecture search / NAS] The performance-aware NAS procedure (described in the architecture search section) provides no details on safeguards against overfitting, such as held-out validation sets distinct from the final evaluation content, cross-validation across domains, or regularization of the search objective. Because the search optimizes directly over perceptual metrics across millions of configurations, this omission is load-bearing for the claimed 20-40% bitrate savings versus other learned codecs.
  2. [Subjective evaluation] The subjective user studies section lacks specification of key experimental controls: number of observers, content selection and diversity, exact comparison protocol (e.g., double-stimulus or forced-choice), and statistical testing for significance. These details are required to substantiate the central 2.3-3x bitrate savings claims against traditional codecs.
  3. [Ablations and evaluation metrics] Exact definitions and implementations of the perceptual metrics used both in the ablations and as the NAS objective are not provided, nor are the training/validation/test data splits. Without these, the support for the reported performance trade-offs cannot be fully assessed.
minor comments (1)
  1. [Abstract] The abstract and runtime claims reference an 'iPhone 17 Pro Max'; if this is a future or hypothetical device, a clarifying footnote would improve precision.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below. Where details were omitted for brevity, we will expand the manuscript to include them, ensuring the claims are fully supported and reproducible.

read point-by-point responses

  1. Referee: [Architecture search / NAS] The performance-aware NAS procedure (described in the architecture search section) provides no details on safeguards against overfitting, such as held-out validation sets distinct from the final evaluation content, cross-validation across domains, or regularization of the search objective. Because the search optimizes directly over perceptual metrics across millions of configurations, this omission is load-bearing for the claimed 20-40% bitrate savings versus other learned codecs.

    Authors: We agree that explicit safeguards must be documented. The NAS was conducted with a held-out validation set of 150 images drawn from the same sources as training data but with no overlap to either training or final test content; the search objective included an explicit runtime regularization term and used early stopping on validation perceptual scores. We also performed the search across multiple random seeds and two domain subsets (natural scenes and high-frequency textures) to assess stability. We will add a dedicated subsection (or appendix) fully describing the NAS protocol, data partitioning, and regularization to make these safeguards transparent. revision: yes

  2. Referee: [Subjective evaluation] The subjective user studies section lacks specification of key experimental controls: number of observers, content selection and diversity, exact comparison protocol (e.g., double-stimulus or forced-choice), and statistical testing for significance. These details are required to substantiate the central 2.3-3x bitrate savings claims against traditional codecs.

    Authors: We will revise the subjective evaluation section to supply all requested controls. The studies used 28 observers screened for normal vision, a double-stimulus continuous quality scale (DSCQS) protocol with randomized side-by-side presentation on calibrated 4K displays, and 45 diverse 12 MP test images spanning natural scenes, portraits, textures, and low-light conditions (none overlapping training data). Statistical significance was assessed via paired t-tests and Wilcoxon signed-rank tests (p < 0.01 for the reported savings). These elements were part of the experimental design; we omitted them for space and will restore them with full protocol description. revision: yes

  3. Referee: [Ablations and evaluation metrics] Exact definitions and implementations of the perceptual metrics used both in the ablations and as the NAS objective are not provided, nor are the training/validation/test data splits. Without these, the support for the reported performance trade-offs cannot be fully assessed.

    Authors: We acknowledge the need for precise metric definitions and splits. LPIPS (AlexNet backbone) and DISTS were used exactly as released by their authors; the NAS objective was a weighted combination of these plus a differentiable runtime proxy. Training used ~2800 images from DIV2K + Flickr2K + internal mobile captures, validation used 120 held-out images from the same pool, and testing used Kodak, CLIC professional, and a separate 50-image 12 MP mobile set. We will insert a concise “Experimental Setup” subsection (or appendix) listing exact metric implementations (with code references), loss weights, and the full data partitioning to enable direct assessment of the ablations and NAS results. revision: yes

Circularity Check

0 steps flagged

No significant circularity in empirical architecture search and evaluations

full rationale

The paper is an empirical study focused on ablations of modeling choices, performance-aware neural architecture search over millions of backbone configurations to optimize perceptual metrics under runtime constraints, and final validation via subjective user studies. No mathematical derivations, first-principles predictions, or results are claimed that reduce to inputs by construction (e.g., no self-definitional equations, fitted parameters renamed as predictions, or load-bearing self-citations). The bitrate savings claims rest on independent subjective evaluations rather than being tautological with the NAS objective. This is a standard non-circular empirical finding.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no specific free parameters, axioms, or invented entities are described. Learned compression models typically involve many training hyperparameters and architectural choices fitted to data.

pith-pipeline@v0.9.0 · 5552 in / 1046 out tokens · 68326 ms · 2026-05-08T16:57:25.229013+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 9 canonical work pages · 2 internal anchors

  1. [1]

    End-to-end optimized image compression.arXiv preprint arXiv:1611.01704, 2016

    J. Ballé, V . Laparra, and E. P. Simoncelli, “End-to- end optimized image compression,”arXiv preprint arXiv:1611.01704, 2016

    work page arXiv 2016

  2. [2]

    Real-time adaptive image compression,

    O. Rippel and L. Bourdev, “Real-time adaptive image compression,” inProceedings of the 34th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, D. Precup and Y . W. Teh, Eds., vol. 70. International Convention Centre, Sydney, Australia: PMLR, 06–11 Aug 2017, pp. 2922– 2930

    2017

  3. [3]

    Variational image compression with a scale hyperprior,

    J. Ballé, D. Minnen, S. Singh, S. J. Hwang, and N. Johnston, “Variational image compression with a scale hyperprior,” inInternational Conference on Learning Representations, 2018. [Online]. Available: https://openreview.net/forum?id=rkcQFMZRb

    2018

  4. [4]

    Integer net- works for data compression with latent-variable mod- els,

    J. Ballé, N. Johnston, and D. Minnen, “Integer net- works for data compression with latent-variable mod- els,” inInternational Conference on Learning Repre- sentations, 2018

    2018

  5. [5]

    Elf-vc: Efficient learned flexible-rate video coding,

    O. Rippel, A. G. Anderson, K. Tatwawadi, S. Nair, C. Lytle, and L. Bourdev, “Elf-vc: Efficient learned flexible-rate video coding,” inProceedings of the IEEE/CVF International Conference on Computer Vi- sion, 2021, pp. 14 479–14 488

    2021

  6. [6]

    Towards real-time neural video codec for cross-platform application using calibration informa- tion,

    K. Tian, Y . Guan, J. Xiang, J. Zhang, X. Han, and W. Yang, “Towards real-time neural video codec for cross-platform application using calibration informa- tion,” inProceedings of the 31st ACM International Conference on Multimedia, 2023, pp. 7961–7970

    2023

  7. [7]

    Towards reproducible learning-based compression,

    J. Pang, M. A. Lodhi, J. Ahn, Y . Huang, and D. Tian, “Towards reproducible learning-based compression,” in 2024 IEEE 26th International Workshop on Multimedia Signal Processing (MMSP). IEEE, 2024, pp. 1–6

    2024

  8. [8]

    Towards practical real-time neural video compression,

    Z. Jia, B. Li, J. Li, W. Xie, L. Qi, H. Li, and Y . Lu, “Towards practical real-time neural video compression,” inIEEE/CVF Conference on Computer Vision and Pat- tern Recognition, CVPR 2025, Nashville, TN, USA, June 11-25, 2025

    2025

  9. [9]

    JPEG AI Reference Software,

    “JPEG AI Reference Software,” https://gitlab.com/ wg1/jpeg-ai/jpeg-ai-reference-software, 2025

    2025

  10. [10]

    High-fidelity generative image com- pression,

    F. Mentzer, G. D. Toderici, M. Tschannen, and E. Agustsson, “High-fidelity generative image com- pression,”Advances in neural information processing systems, vol. 33, pp. 11 913–11 924, 2020

    2020

  11. [11]

    Lossy image compression with conditional diffusion models,

    R. Yang and S. Mandt, “Lossy image compression with conditional diffusion models,”Advances in Neural Information Processing Systems, vol. 36, pp. 64 971– 64 995, 2023

    2023

  12. [12]

    Good, cheap, and fast: Overfitted image compression with wasserstein distortion,

    J. Ballé, L. Versari, E. Dupont, H. Kim, and M. Bauer, “Good, cheap, and fast: Overfitted image compression with wasserstein distortion,” inProceedings of the Com- puter Vision and Pattern Recognition Conference, 2025, pp. 23 259–23 268

    2025

  13. [13]

    Po-elic: Perception- oriented efficient learned image coding,

    D. He, Z. Yang, H. Yu, T. Xu, J. Luo, Y . Chen, C. Gao, X. Shi, H. Qin, and Y . Wang, “Po-elic: Perception- oriented efficient learned image coding,” 2022. [Online]. Available: https://arxiv.org/abs/2205.14501

    work page arXiv 2022

  14. [14]

    libbpg: BPG (Better Portable Graphics) image library,

    F. Bellard, “libbpg: BPG (Better Portable Graphics) image library,” http://bellard.org/bpg/libbpg-0.9.5.tar. gz, 2015, version 0.9.5, released 2015-01-11

    2015

  15. [15]

    VVCSoft- ware_VTM: VVC VTM Reference Software,

    Joint Video Experts Team (JVET), “VVCSoft- ware_VTM: VVC VTM Reference Software,” https: //vcgit.hhi.fraunhofer.de/jvet/VVCSoftware_VTM/- /releases/VTM-23.11, 2025, version 23.11, released 2025-07-03

    2025

  16. [16]

    AOM: A V1 codec library, version 3.12.1,

    Alliance for Open Media, “AOM: A V1 codec library, version 3.12.1,” Alliance for Open Media, Git repository, April 2025, tag: v3.12.1, commit 10aece4, tagged fc5cf6a. [Online]. Available: https://aomedia.googlesource.com/aom/+/refs/tags/ v3.12.1

    2025

  17. [17]

    Enhanced Compression Model (ECM) Reference Software,

    Joint Video Experts Team (JVET), “Enhanced Compression Model (ECM) Reference Software,” Fraunhofer HHI, GitLab repository, September 2025, branch: master, commit dcc311af, accessed 2025-09-25. [Online]. Available: https: //vcgit.hhi.fraunhofer.de/ecm/ECM

    2025

  18. [18]

    A VM: A V2 codec research anchor, version research-v11.0.0,

    Alliance for Open Media, “A VM: A V2 codec research anchor, version research-v11.0.0,” Alliance for Open Media, GitLab repository, September 2025, tag: research-v11.0.0, commit 3a5da21a. [On- line]. Available: https://gitlab.com/AOMediaCodec/ avm/-/tags/research-v11.0.0

    2025

  19. [19]

    Michael S

    E. Agustsson, M. Tschannen, F. Mentzer, R. Timofte, and L. Van Gool, “Generative adversarial networks for extreme learned image compression,”arXiv preprint arXiv:1804.02958, 2018

    work page arXiv 2018

  20. [20]

    Multi-realism image compression with a conditional generator,

    E. Agustsson, D. Minnen, G. Toderici, and F. Mentzer, “Multi-realism image compression with a conditional generator,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 2023, pp. 22 324–22 333

    2023

  21. [21]

    Generative la- tent coding for ultra-low bitrate image compression,

    Z. Jia, J. Li, B. Li, H. Li, and Y . Lu, “Generative la- tent coding for ultra-low bitrate image compression,” inProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2024, pp. 26 088– 26 098

    2024

  22. [22]

    Bridging the gap between diffusion mod- els and universal quantization for image compression,

    L. Relic, R. Azevedo, Y . Zhang, M. Gross, and C. Schroers, “Bridging the gap between diffusion mod- els and universal quantization for image compression,” inMachine Learning and Compression Workshop@ NeurIPS 2024. OpenReview, 2024

    2024

  23. [23]

    Available: https://orange-opensource.github

    [Online]. Available: https://orange-opensource.github. io/Cool-Chic/

  24. [24]

    Joint Autoregressive and Hierarchical Priors for Learned Image Compression

    D. Minnen, J. Ballé, and G. Toderici, “Joint autoregres- sive and hierarchical priors for learned image compres- sion,”arXiv preprint arXiv:1809.02736, 2018

    work page Pith review arXiv 2018

  25. [25]

    Checkerboard context model for efficient learned im- age compression,

    D. He, Y . Zheng, B. Sun, Y . Wang, and H. Qin, “Checkerboard context model for efficient learned im- age compression,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2021, pp. 14 771–14 780

    2021

  26. [26]

    Elic: Efficient learned image compression with un- evenly grouped space-channel contextual adaptive cod- ing,

    D. He, Z. Yang, W. Peng, R. Ma, H. Qin, and Y . Wang, “Elic: Efficient learned image compression with un- evenly grouped space-channel contextual adaptive cod- ing,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5718–5727

    2022

  27. [27]

    Neural video compression with feature modulation,

    J. Li, B. Li, and Y . Lu, “Neural video compression with feature modulation,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 17-21, 2024, 2024

    2024

  28. [28]

    C3: High-performance and low- complexity neural compression from a single image or video,

    H. Kim, M. Bauer, L. Theis, J. R. Schwarz, and E. Dupont, “C3: High-performance and low- complexity neural compression from a single image or video,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 9347–9358

    2024

  29. [29]

    Learning a deep vector quantization network for im- age compression,

    X. Lu, H. Wang, W. Dong, F. Wu, Z. Zheng, and G. Shi, “Learning a deep vector quantization network for im- age compression,”IEEE Access, vol. 7, pp. 118 815– 118 825, 2019

    2019

  30. [30]

    Ntire 2017 challenge on single image super-resolution: Dataset and study,

    E. Agustsson and R. Timofte, “Ntire 2017 challenge on single image super-resolution: Dataset and study,” in The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, July 2017

    2017

  31. [31]

    Available: https://clic2025.compression.cc/ tasks/#image

    [Online]. Available: https://clic2025.compression.cc/ tasks/#image

  32. [32]

    Ntire 2017 challenge on single image super-resolution: Methods and results,

    R. Timofte, E. Agustsson, L. V . Gool, M.-H. Yang, L. Zhang, B. Lim, S. Son, H. Kim, S. Nah, K. M. Lee, X. Wang, Y . Tian, K. Yu, Y . Zhang, S. Wu, C. Dong, L. Lin, Y . Qiao, C. C. Loy, W. Bae, J. Yoo, Y . Han, J. C. Ye, J.-S. Choi, M. Kim, Y . Fan, J. Yu, W. Han, D. Liu, H. Yu, Z. Wang, H. Shi, X. Wang, T. S. Huang, Y . Chen, K. Zhang, W. Zuo, Z. Tang, L...

    2017

  33. [33]

    Adam: A Method for Stochastic Optimization

    D. Kingma and J. Ba, “Adam: A method for stochastic optimization,”arXiv preprint arXiv:1412.6980, 2014

    work page internal anchor Pith review arXiv 2014

  34. [34]

    Mobilenetv2: Inverted residuals and linear bottlenecks,

    M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.- C. Chen, “Mobilenetv2: Inverted residuals and linear bottlenecks,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 4510–4520

    2018

  35. [35]

    Channel-wise autoregres- sive entropy models for learned image compression,

    D. Minnen and S. Singh, “Channel-wise autoregres- sive entropy models for learned image compression,” in2020 IEEE International Conference on Image Pro- cessing (ICIP). IEEE, 2020, pp. 3339–3343

    2020

  36. [36]

    Cosmos World Foundation Model Platform for Physical AI

    NVIDIA: N. Agarwal, A. Ali, M. Bala, Y . Balaji, E. Barker, T. Cai, P. Chattopadhyay, Y . Chen, Y . Cui, Y . Ding, D. Dworakowski, J. Fan, M. Fenzi, F. Ferroni, S. Fidler, D. Fox, S. Ge, Y . Ge, J. Gu, S. Gururani, E. He, J. Huang, J. Huffman, P. Jannaty, J. Jin, S. W. Kim, G. Klár, G. Lam, S. Lan, L. Leal-Taixe, A. Li, Z. Li, C.-H. Lin, T.-Y . Lin, H. Li...

    work page internal anchor Pith review arXiv 2025

  37. [37]

    The unreasonable effectiveness of deep fea- tures as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep fea- tures as a perceptual metric,” inProceedings of the IEEE conference on computer vision and pattern recog- nition, 2018, pp. 586–595

    2018

  38. [38]

    Multi- scale structural similarity for image quality assessment,

    Z. Wang, E. P. Simoncelli, and A. C. Bovik, “Multi- scale structural similarity for image quality assessment,” inSignals, Systems and Computers, 2004., vol. 2. Ieee, 2003, pp. 1398–1402

    2004

  39. [39]

    Charac- ter region awareness for text detection,

    Y . Baek, B. Lee, D. Han, S. Yun, and H. Lee, “Charac- ter region awareness for text detection,” inProceedings of the IEEE Conference on Computer Vision and Pat- tern Recognition, 2019, pp. 9365–9374

    2019

  40. [40]

    Image- to-image translation with conditional adversarial net- works,

    P. Isola, J.-Y . Zhu, T. Zhou, and A. A. Efros, “Image- to-image translation with conditional adversarial net- works,” inProceedings of the IEEE conference on com- puter vision and pattern recognition, 2017, pp. 1125– 1134

    2017

  41. [41]

    Mlic++: Linear complexity multi-reference entropy modeling for learned image compression,

    W. Jiang, J. Yang, Y . Zhai, F. Gao, and R. Wang, “Mlic++: Linear complexity multi-reference entropy modeling for learned image compression,”arXiv preprint arXiv:2307.15421, 2023

    work page arXiv 2023

  42. [42]

    Learned image compres- sion with mixed transformer-cnn architectures,

    J. Liu, H. Sun, and J. Katto, “Learned image compres- sion with mixed transformer-cnn architectures,” inPro- ceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, 2023, pp. 14 388–14 397

    2023

  43. [43]

    Rethinking fid: To- wards a better evaluation metric for image generation,

    S. Jayasumana, S. Ramalingam, A. Veit, D. Glasner, A. Chakrabarti, and S. Kumar, “Rethinking fid: To- wards a better evaluation metric for image generation,” inProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2024, pp. 9307– 9315

    2024

  44. [44]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” Advances in neural information processing systems, vol. 30, 2017

    2017

  45. [45]

    Available: https://www.mabyduck.com

    [Online]. Available: https://www.mabyduck.com

  46. [46]

    A crowdsourcing approach to video quality assessment,

    B. Naderi and R. Cutler, “A crowdsourcing approach to video quality assessment,” 2023. [Online]. Available: https://arxiv.org/abs/2204.06784

    work page arXiv 2023

  47. [47]

    Efficient bayesian inference for generalized bradley-terry models,

    F. Caron and A. Doucet, “Efficient bayesian inference for generalized bradley-terry models,” 2010. [Online]. Available: https://arxiv.org/abs/1011.1761

    work page arXiv 2010

  48. [48]

    Available: https://docs.mabyduck.com/ experiments/metrics/elo

    [Online]. Available: https://docs.mabyduck.com/ experiments/metrics/elo

  49. [49]

    Available: https://docs.mabyduck.com/ experiments/strategies

    [Online]. Available: https://docs.mabyduck.com/ experiments/strategies

  50. [50]

    <src>" <pad_option>’-pix_fmt yuv420p

    S. Ishihara,Tests for color-blindness. Tokyo, Hongo Harukicho: Handaya, 1917. A. Additional evaluations Figure 8 presents curves for additional metrics. Although the perceptual codecs PICO, HiFiC, C3-WD and CDC sub- stantially outperform the non-perceptual codecs based on human ratings and the perceptually-oriented objective met- rics, they do not perform...

    1917