pith. sign in

arxiv: 2605.16817 · v1 · pith:3XU2JJFLnew · submitted 2026-05-16 · 📡 eess.IV · cs.CV

Adaptive Fused Prior Transfer for Controllable Generative Image Compression

Yifei Pei , Ying Liu , Nam Ling This is my paper

Pith reviewed 2026-05-19 19:53 UTC · model grok-4.3

classification 📡 eess.IV cs.CV
keywords learned image compressiongenerative priorscontrollable codecsfused prior transferlow-bitrate reconstructionperceptual quality
0
0 comments X

The pith

AFP-GIC transfers fused priors from a frozen model so the decoder can predict them from compressed data and controls alone.

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

The paper introduces AFP-GIC, a controllable generative image codec that avoids transmitting reconstruction priors by letting the decoder predict an adaptive fused prior. Encoder features derived from a frozen pretrained AdaCode model shape the latent during compression. The decoder then reconstructs a matching fused prior using only the received bitstream and chosen control variables, guiding detail recovery at very low rates. This design is shown to contain single-codebook methods as special cases and to produce lower decoder latency and smaller model size than prior controllable codecs while matching or exceeding their perceptual quality on standard test sets.

Core claim

AFP-GIC transfers an adaptive fused prior from a frozen pretrained AdaCode model. Encoder-side fused-prior features guide latent formation, while the decoder predicts a compatible fused prior from the compressed representation and selected control variables, enabling prior-guided reconstruction without transmitting the fused prior itself. A motivating analysis relates decoder-side fused-prior alignment to a reconstruction-error upper bound and shows that the fused-prior family contains single-codebook choices as special cases.

What carries the argument

Decoder-side prediction of an adaptive fused prior from the compressed latent and control variables, guided by encoder features taken from a frozen pretrained model.

If this is right

  • Decoder latency drops 18.1 percent relative to DC-VIC.
  • Total parameter count falls by 31.10 million parameters, or 20.5 percent.
  • PSNR remains competitive with existing controllable codecs on Kodak, CLIC2020, and DIV2K.
  • NIQE scores and visual quality at very low bitrates improve over prior single-codebook designs.

Where Pith is reading between the lines

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

  • The same prior-prediction strategy could reduce transmission overhead in other generative tasks that currently send explicit codebook indices.
  • Because the fused prior contains single-codebook priors as special cases, the method supplies a unified way to trade off between reconstruction fidelity and controllability without changing the model architecture.
  • If the prediction step proves robust across datasets, real-time mobile decoders could adopt fused-prior guidance without increasing on-device memory.

Load-bearing premise

The decoder can reliably predict a compatible fused prior from only the compressed representation and selected control variables.

What would settle it

Measure whether the reconstruction error on held-out images exceeds the paper's stated upper bound when the decoder must predict the fused prior from bitstreams that were encoded with deliberately mismatched control variables.

Figures

Figures reproduced from arXiv: 2605.16817 by Nam Ling, Yifei Pei, Ying Liu.

Figure 1
Figure 1. Figure 1: Overview of the proposed Adaptive Fused Prior Transfer for Controllable Generative Image Compression (AFP-GIC). [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Frozen AdaCode prior extractor. The input image is encoded and quantized by multiple codebooks. The resulting [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Quantitative comparison with learned generative codecs on (a) Kodak, (b) CLIC2020, and (c) DIV2K. “ [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Quantitative comparison of AFP-GIC, VVC Intra, and BPG on CLIC2020. “ [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIGURE 5 [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIGURE 6 [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIGURE 7 [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison between the representative single-prior baseline C2 and AFP-GIC at close bitrates for a Kodak image. [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Effects of βrate and βprior in AFP-GIC. (a) The selected βrate and βprior values across the five target bitrates. (b) With βrate fixed for each target bitrate, PSNR change relative to βprior = 0 is plotted over eight βprior values in [0, 3.5]. Impact of Adversarial Supervision [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Visual effect of adversarial supervision at 0.0833 bpp [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
read the original abstract

Learned image compression has achieved competitive rate-distortion performance, but very-low-bitrate reconstruction remains difficult because the transmitted representation often cannot preserve fine textures and local structures. Perceptual and generative codecs address this problem by using learned reconstruction priors, and controllable codecs allow one model to cover different bitrate and reconstruction preferences. However, controllability alone does not resolve the decoder-side reconstruction-prior problem: under severe bit constraints, the decoder must infer missing details from limited transmitted information, while existing codebook-based controllable designs generally rely on single-codebook token-based priors. This paper proposes Adaptive Fused Prior Transfer for Controllable Generative Image Compression (AFP-GIC), a controllable codec that transfers an adaptive fused prior from a frozen pretrained AdaCode model. Encoder-side fused-prior features guide latent formation, while the decoder predicts a compatible fused prior from the compressed representation and selected control variables, enabling prior-guided reconstruction without transmitting the fused prior itself. A motivating analysis relates decoder-side fused-prior alignment to a reconstruction-error upper bound and shows that the fused-prior family contains single-codebook choices as special cases. Under the unified benchmark, AFP-GIC reduces decoder latency by 18.1% and the overall parameter count by 31.10 million (20.5%) relative to DC-VIC. Experiments on Kodak, CLIC2020, and DIV2K show competitive PSNR, with the clearest perceptual gains in NIQE scores and very-low-bitrate visual comparisons.

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 / 3 minor

Summary. The manuscript introduces AFP-GIC, a controllable generative image compression codec that transfers an adaptive fused prior from a frozen pretrained AdaCode model. Encoder-side fused-prior features guide latent formation, while the decoder predicts a compatible fused prior from the compressed representation and control variables to enable prior-guided reconstruction without transmitting the prior. A motivating analysis relates decoder-side alignment to a reconstruction-error upper bound and shows that the fused-prior family subsumes single-codebook priors as special cases. Experiments on Kodak, CLIC2020, and DIV2K report competitive PSNR, superior NIQE scores at very low bitrates, an 18.1% decoder latency reduction, and a 20.5% (31.10 million) overall parameter reduction relative to DC-VIC.

Significance. If the decoder-side prior prediction holds, the method provides a practical route to controllable generative compression with reduced transmission overhead and model size while improving perceptual quality at low rates. The latency and parameter savings are concrete and the NIQE/visual gains at very low bitrates address a known weakness of learned codecs. The generalization argument and alignment-to-error-bound analysis supply theoretical support that elevates the work beyond pure empirical tuning.

major comments (3)
  1. [Motivating Analysis and Method Overview] Motivating Analysis and Method Overview: the central claim that decoder-side fused-prior alignment yields the observed perceptual gains rests on the assumption that the learned predictor achieves sufficient alignment in the low-bitrate regime. No explicit quantification of achieved alignment (cosine similarity, feature-space distance, or similar) between encoder and decoder priors is reported, nor is there an ablation isolating the effect of misalignment on PSNR/NIQE. Without these, the upper-bound relation does not directly explain the reported improvements.
  2. [Experiments] Experiments section: the 18.1% latency and 20.5% parameter reductions, as well as the NIQE and visual claims, are presented without error bars, dataset split details, or full protocol (training/validation/test divisions, number of runs, random seeds). This information is required to assess whether the gains over DC-VIC are statistically reliable and reproducible.
  3. [Method Overview] Method Overview: the decoder must infer a compatible fused prior from compressed latents plus control variables alone. No ablation is described that varies the quality of this prediction (e.g., by injecting controlled misalignment) and measures impact on reconstruction metrics, leaving the weakest assumption untested.
minor comments (3)
  1. Add error bars or confidence intervals to all quantitative tables and plots.
  2. Clarify the exact form of the control variables and how they are concatenated with the compressed representation before prior prediction.
  3. Include a diagram showing the encoder/decoder prior paths and the frozen AdaCode component to improve readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major comment below and indicate the revisions planned for the next manuscript version.

read point-by-point responses

  1. Referee: [Motivating Analysis and Method Overview] Motivating Analysis and Method Overview: the central claim that decoder-side fused-prior alignment yields the observed perceptual gains rests on the assumption that the learned predictor achieves sufficient alignment in the low-bitrate regime. No explicit quantification of achieved alignment (cosine similarity, feature-space distance, or similar) between encoder and decoder priors is reported, nor is there an ablation isolating the effect of misalignment on PSNR/NIQE. Without these, the upper-bound relation does not directly explain the reported improvements.

    Authors: We agree that direct quantification of alignment would strengthen the link between the motivating analysis and the reported gains. In the revised manuscript we add explicit measurements of cosine similarity and feature-space distance between the encoder-side and decoder-side fused priors, reported across bitrate regimes. We have also performed and included an ablation that introduces controlled misalignment and measures its effect on PSNR and NIQE; the results are now presented alongside the upper-bound derivation. revision: yes

  2. Referee: [Experiments] Experiments section: the 18.1% latency and 20.5% parameter reductions, as well as the NIQE and visual claims, are presented without error bars, dataset split details, or full protocol (training/validation/test divisions, number of runs, random seeds). This information is required to assess whether the gains over DC-VIC are statistically reliable and reproducible.

    Authors: We acknowledge that the current presentation lacks the statistical and procedural details needed for reproducibility assessment. The revised manuscript now reports error bars computed over multiple independent runs with documented random seeds, provides the exact training/validation/test divisions used for Kodak, CLIC2020 and DIV2K, and includes the complete experimental protocol (hyper-parameters, number of runs, and seeds). revision: yes

  3. Referee: [Method Overview] Method Overview: the decoder must infer a compatible fused prior from compressed latents plus control variables alone. No ablation is described that varies the quality of this prediction (e.g., by injecting controlled misalignment) and measures impact on reconstruction metrics, leaving the weakest assumption untested.

    Authors: We recognize the value of directly testing the decoder-side prior prediction assumption. We have added an ablation study in which controlled misalignment is injected at varying degrees into the predicted fused prior; the resulting changes in reconstruction metrics are reported in the revised Method Overview section together with supporting figures and analysis. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper introduces the AFP-GIC method, including encoder-side guidance from fused priors of a frozen AdaCode model and decoder-side prediction of compatible priors from compressed latents plus controls. The motivating analysis links decoder alignment to a reconstruction error upper bound and notes that the fused-prior family subsumes single-codebook priors as special cases. No equations, self-citations, or fitted parameters are shown in the text that reduce these relations or the reported latency/parameter reductions and NIQE gains to tautological inputs by construction. Performance claims rest on empirical results across Kodak, CLIC2020, and DIV2K rather than self-referential definitions or renamed fits, rendering the chain self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review prevents identification of concrete free parameters or axioms; the method appears to rely on standard assumptions of learned compression and pretrained generative priors.

pith-pipeline@v0.9.0 · 5795 in / 1102 out tokens · 47186 ms · 2026-05-19T19:53:30.848480+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

53 extracted references · 53 canonical work pages · 1 internal anchor

  1. [1]

    Nonlinear transform coding,

    J. Ballé, P. A. Chou, D. Minnen, S. Singh, N. Johnston, E. Agustsson, S. J. Hwang, and G. Toderici, “Nonlinear transform coding,”IEEE J. Sel. Topics Signal Process., vol. 15, no. 2, pp. 339–353, 2021

    work page 2021

  2. [2]

    End-to-end optimized image compression,

    J. Ballé, V . Laparra, and E. P. Simoncelli, “End-to-end optimized image compression,” inInt. Conf. Learn. Rep- resent. (ICLR), 2017

    work page 2017

  3. [3]

    Variational image compression with a scale hyper- prior,

    J. Ballé, D. Minnen, S. Singh, S. J. Hwang, and N. John- ston, “Variational image compression with a scale hyper- prior,” inInt. Conf. Learn. Represent. (ICLR), 2018

    work page 2018

  4. [4]

    Joint autore- gressive and hierarchical priors for learned image com- pression,

    D. Minnen, J. Ballé, and G. D. Toderici, “Joint autore- gressive and hierarchical priors for learned image com- pression,” inAdv. Neural Inf. Process. Syst., 2018, pp. 10 771–10 780

    work page 2018

  5. [5]

    Conditional probability models for deep im- age compression,

    F. Mentzer, E. Agustsson, M. Tschannen, R. Timofte, and L. V . Gool, “Conditional probability models for deep im- age compression,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2018, pp. 4394–4402. Preprint– AdaptiveFusedPriorTransfer forControllableGenerativeImageCompression 1 18

    work page 2018

  6. [6]

    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,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022, pp. 5708–5717

    work page 2022

  7. [7]

    Learning end-to-end lossy image compression: A benchmark,

    Y . Hu, W. Yang, Z. Ma, and J. Liu, “Learning end-to-end lossy image compression: A benchmark,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 44, no. 8, pp. 4194–4211, 2022

    work page 2022

  8. [8]

    Mean squared error: Love it or leave it? A new look at signal fidelity measures,

    Z. Wang and A. C. Bovik, “Mean squared error: Love it or leave it? A new look at signal fidelity measures,”IEEE Signal Process. Mag., vol. 26, no. 1, pp. 98–117, 2009

    work page 2009

  9. [9]

    The perception-distortion trade- off,

    Y . Blau and T. Michaeli, “The perception-distortion trade- off,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2018, pp. 6228–6237

    work page 2018

  10. [10]

    Rethinking lossy compres- sion: The rate-distortion-perception tradeoff,

    Y . Blau and T. Michaeli, “Rethinking lossy compres- sion: The rate-distortion-perception tradeoff,” inProc. Int. Conf. Mach. Learn. (ICML), ser. Proceedings of Machine Learning Research, vol. 97, 2019, pp. 675–685

    work page 2019

  11. [11]

    High-fidelity generative image compression,

    F. Mentzer, G. D. Toderici, M. Tschannen, and E. Agusts- son, “High-fidelity generative image compression,” in Adv. Neural Inf. Process. Syst., vol. 33, 2020, pp. 11 913– 11 924

    work page 2020

  12. [12]

    Improving statistical fidelity for neural im- age compression with implicit local likelihood models,

    M. J. Muckley, A. El-Nouby, K. Ullrich, H. Jégou, and J. Verbeek, “Improving statistical fidelity for neural im- age compression with implicit local likelihood models,” inProc. Int. Conf. Mach. Learn. (ICML), ser. Proceed- ings of Machine Learning Research, vol. 202, 2023, pp. 25 426–25 443

    work page 2023

  13. [13]

    Towards image compression with perfect realism at ultra-low bitrates,

    M. Careil, M. J. Muckley, J. Verbeek, and S. Lathuil- ière, “Towards image compression with perfect realism at ultra-low bitrates,” inInt. Conf. Learn. Represent. (ICLR), 2024

    work page 2024

  14. [14]

    Lossy image compression with conditional diffusion models,

    R. Yang and S. Mandt, “Lossy image compression with conditional diffusion models,” inAdv. Neural Inf. Process. Syst., vol. 36, 2023, pp. 64 971–64 995

    work page 2023

  15. [15]

    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,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023, pp. 22 324–22 333

    work page 2023

  16. [16]

    Controlling rate, distortion, and realism: Towards a single comprehensive neural image compression model,

    S. Iwai, T. Miyazaki, and S. Omachi, “Controlling rate, distortion, and realism: Towards a single comprehensive neural image compression model,” inProc. IEEE/CVF Winter Conf. Appl. Comput. Vis. (WACV), 2024, pp. 2888– 2897

    work page 2024

  17. [17]

    Once- for-all: Controllable generative image compression with dynamic granularity adaptation,

    A. Li, F. Li, Y . Liu, R. Cong, Y . Zhao, and H. Bai, “Once- for-all: Controllable generative image compression with dynamic granularity adaptation,” inInt. Conf. Learn. Rep- resent. (ICLR), 2025

    work page 2025

  18. [18]

    Dual-conditioned training to exploit pre-trained codebook-based generative model in image compression,

    S. Iwai, T. Miyazaki, and S. Omachi, “Dual-conditioned training to exploit pre-trained codebook-based generative model in image compression,”IEEE Access, vol. 12, pp. 198 184–198 200, 2024

    work page 2024

  19. [19]

    Neu- ral discrete representation learning,

    A. van den Oord, O. Vinyals, and K. Kavukcuoglu, “Neu- ral discrete representation learning,” inAdv. Neural Inf. Process. Syst., 2017, pp. 6306–6315

    work page 2017

  20. [20]

    Taming trans- formers for high-resolution image synthesis,

    P. Esser, R. Rombach, and B. Ommer, “Taming trans- formers for high-resolution image synthesis,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2021, pp. 12 868–12 878

    work page 2021

  21. [21]

    Learning image- adaptive codebooks for class-agnostic image restoration,

    K. Liu, Y . Jiang, I. Choi, and J. Gu, “Learning image- adaptive codebooks for class-agnostic image restoration,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2023, pp. 5350–5360

    work page 2023

  22. [22]

    Learned image compression with discretized Gaussian mixture likelihoods and attention modules,

    Z. Cheng, H. Sun, M. Takeuchi, and J. Katto, “Learned image compression with discretized Gaussian mixture likelihoods and attention modules,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2020, pp. 7936–7945

    work page 2020

  23. [23]

    Deep gen- erative models for distribution-preserving lossy compres- sion,

    M. Tschannen, E. Agustsson, and M. Lucic, “Deep gen- erative models for distribution-preserving lossy compres- sion,” inAdv. Neural Inf. Process. Syst., vol. 31, 2018, pp. 5929–5940

    work page 2018

  24. [24]

    Generative adversarial networks for ex- treme learned image compression,

    E. Agustsson, M. Tschannen, F. Mentzer, R. Timofte, and L. V . Gool, “Generative adversarial networks for ex- treme learned image compression,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2019, pp. 221–231

    work page 2019

  25. [25]

    Generative latent coding for ultra-low bitrate image compression,

    Z. Jia, J. Li, B. Li, H. Li, and Y . Lu, “Generative latent coding for ultra-low bitrate image compression,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024, pp. 26 088–26 098

    work page 2024

  26. [26]

    Lossy image compression with foundation diffusion models,

    L. Relic, R. Azevedo, M. Gross, and C. Schroers, “Lossy image compression with foundation diffusion models,” in Proc. Eur. Conf. Comput. Vis. (ECCV), 2024, pp. 303– 319

    work page 2024

  27. [27]

    Variable rate deep image compression with a conditional autoencoder,

    Y . Choi, M. El-Khamy, and J. Lee, “Variable rate deep image compression with a conditional autoencoder,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2019, pp. 3146–3154

    work page 2019

  28. [28]

    Asymmetric gained deep image compression with con- tinuous rate adaptation,

    Z. Cui, J. Wang, S. Gao, T. Guo, Y . Feng, and B. Bai, “Asymmetric gained deep image compression with con- tinuous rate adaptation,” inProc. IEEE/CVF Conf. Com- put. Vis. Pattern Recognit. (CVPR), 2021, pp. 10 527– 10 536

    work page 2021

  29. [29]

    Variable-rate deep im- age compression through spatially-adaptive feature trans- form,

    M. Song, J. Choi, and B. Han, “Variable-rate deep im- age compression through spatially-adaptive feature trans- form,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2021, pp. 2360–2369

    work page 2021

  30. [30]

    Fidelity- controllable extreme image compression with generative adversarial networks,

    S. Iwai, T. Miyazaki, Y . Sugaya, and S. Omachi, “Fidelity- controllable extreme image compression with generative adversarial networks,” inProc. Int. Conf. Pattern Recog- nit. (ICPR), 2021, pp. 8235–8242

    work page 2021

  31. [31]

    Generat- ing diverse high-fidelity images with VQ-V AE-2,

    A. Razavi, A. van den Oord, and O. Vinyals, “Generat- ing diverse high-fidelity images with VQ-V AE-2,” inAdv. Neural Inf. Process. Syst., 2019, pp. 14 837–14 847

    work page 2019

  32. [32]

    Extreme image compression using fine-tuned VQGANs,

    Q. Mao, T. Yang, Y . Zhang, Z. Wang, M. Wang, S. Wang, L. Jin, and S. Ma, “Extreme image compression using fine-tuned VQGANs,” inProc. Data Compression Conf. (DCC), 2024, pp. 203–212

    work page 2024

  33. [33]

    Recovering realistic texture in image super-resolution by deep spatial feature transform,

    X. Wang, K. Yu, C. Dong, and C. C. Loy, “Recovering realistic texture in image super-resolution by deep spatial feature transform,” inProc. IEEE Conf. Comput. Vis. Pat- tern Recognit. (CVPR), 2018, pp. 606–615. Preprint– AdaptiveFusedPriorTransfer forControllableGenerativeImageCompression 1 19

    work page 2018

  34. [34]

    Group normalization,

    Y . Wu and K. He, “Group normalization,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2018, pp. 3–19

    work page 2018

  35. [35]

    Sigmoid-weighted linear units for neural network function approximation in reinforcement learning,

    S. Elfwing, E. Uchibe, and K. Doya, “Sigmoid-weighted linear units for neural network function approximation in reinforcement learning,”Neural Networks, vol. 107, pp. 3–11, 2018

    work page 2018

  36. [36]

    Fourier features let networks learn high-frequency functions in low-dimensional domains,

    M. Tancik, P. P. Srinivasan, B. Mildenhall, S. Fridovich- Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. T. Barron, and R. Ng, “Fourier features let networks learn high-frequency functions in low-dimensional domains,” inAdv. Neural Inf. Process. Syst., vol. 33, 2020, pp. 7537– 7547

    work page 2020

  37. [37]

    Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation

    Y . Bengio, N. Léonard, and A. Courville, “Estimating or propagating gradients through stochastic neurons for con- ditional computation,”arXiv preprint arXiv:1308.3432, 2013

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  38. [38]

    Image-to- image translation with conditional adversarial networks,

    P. Isola, J.-Y . Zhu, T. Zhou, and A. A. Efros, “Image-to- image translation with conditional adversarial networks,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2017, pp. 5967–5976

    work page 2017

  39. [39]

    The Open Images Dataset V4,

    A. Kuznetsova, H. Rom, N. Alldrin, J. R. R. Uijlings, I. Krasin, J. Pont-Tuset, S. Kamali, S. Popov, M. Mal- loci, A. Kolesnikov, T. Duerig, and V . Ferrari, “The Open Images Dataset V4,”Int. J. Comput. Vis., vol. 128, no. 7, pp. 1956–1981, 2020

    work page 1956

  40. [40]

    Kodak lossless true color image suite,

    “Kodak lossless true color image suite,” Official dataset page, accessed: May 2, 2026. [Online]. Available: https://r0k.us/graphics/kodak/

    work page 2026

  41. [41]

    Work- shop and challenge on learned image compression (CLIC2020),

    G. Toderici, W. Shi, R. Timofte, L. Theis, J. Ballé, E. Agustsson, N. Johnston, and F. Mentzer, “Work- shop and challenge on learned image compression (CLIC2020),” CVPR, 2020, accessed: May 2, 2026. [Online]. Available: https://archive.compression.cc/2020/

    work page 2020

  42. [42]

    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 Proc. IEEE Conf. Comput. Vis. Pattern Recognit. Work- shops (CVPRW), 2017, pp. 1122–1131

    work page 2017

  43. [43]

    Overview of the versatile video coding (VVC) standard and its applications,

    B. Bross, Y .-K. Wang, Y . Ye, S. Liu, J. Chen, G. J. Sul- livan, and J.-R. Ohm, “Overview of the versatile video coding (VVC) standard and its applications,”IEEE Trans. Circuits Syst. Video Technol., vol. 31, no. 10, pp. 3736– 3764, 2021

    work page 2021

  44. [44]

    BPG image format,

    F. Bellard, “BPG image format,” Official technical specification/project page, 2014, version 0.9.8; accessed: May 2, 2026. [Online]. Available: https://bellard.org/bpg/

    work page 2014

  45. [45]

    VVenC: An open and optimized VVC encoder implementation,

    A. Wieckowski, J. Brandenburg, T. Hinz, C. Bart- nik, V . George, G. Hege, C. Helmrich, A. Henkel, C. Lehmann, C. Stoffers, I. Zupancic, B. Bross, and D. Marpe, “VVenC: An open and optimized VVC encoder implementation,” inProc. IEEE International Conference on Multimedia Expo Workshops (ICMEW), 2021, pp. 1–2

    work page 2021

  46. [46]

    Towards a live soft- ware decoder implementation for the upcoming versatile video coding (VVC) codec,

    A. Wieckowski, G. Hege, C. Bartnik, C. Lehmann, C. Stoffers, B. Bross, and D. Marpe, “Towards a live soft- ware decoder implementation for the upcoming versatile video coding (VVC) codec,” inProc. IEEE International Conference on Image Processing (ICIP), 2020, pp. 3124– 3128

    work page 2020

  47. [47]

    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,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2018, pp. 586– 595

    work page 2018

  48. [48]

    Image quality assessment: Unifying structure and texture simi- larity,

    K. Ding, K. Ma, S. Wang, and E. P. Simoncelli, “Image quality assessment: Unifying structure and texture simi- larity,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 44, no. 5, pp. 2567–2581, 2022

    work page 2022

  49. [49]

    Making a “completely blind

    A. Mittal, R. Soundararajan, and A. C. Bovik, “Making a “completely blind” image quality analyzer,”IEEE Signal Process. Lett., vol. 20, no. 3, pp. 209–212, 2013

    work page 2013

  50. [50]

    Multiscale structural similarity for image quality assessment,

    Z. Wang, E. P. Simoncelli, and A. C. Bovik, “Multiscale structural similarity for image quality assessment,” in Proc. 37th Asilomar Conf. Signals, Syst. Comput., vol. 2, 2003, pp. 1398–1402

    work page 2003

  51. [51]

    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,” inAdv. Neural Inf. Process. Syst., 2017, pp. 6626–6637

    work page 2017

  52. [52]

    Calculation of average PSNR differ- ences between RD-curves,

    G. Bjøntegaard, “Calculation of average PSNR differ- ences between RD-curves,” ITU-T VCEG document VCEG-M33, April 2001, accessed: May 2, 2026. [Online]. Available: https://www.itu.int/wftp3/av-arch/ video-site/0104_Aus/VCEG-M33.doc

    work page 2001

  53. [53]

    Adam: A method for stochas- tic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochas- tic optimization,” inInt. Conf. Learn. Represent. (ICLR), 2015

    work page 2015