Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

StableCodec: Taming One-Step Diffusion for Extreme Image Compression

T0 review · 2 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read StableCodec claims a single denoising step can restore images compressed to 0.005 bits per pixel.

desk verdict One-step diffusion codec is a real speed win and the evaluation is unusually thorough, but the color-correction side-channel in the supplement needs an ablation before the SOTA margins are taken at face value. read the letter →

arxiv 2506.21977 v1 pith:TQ2QPTFU submitted 2025-06-27 eess.IV cs.CV

Tianyu Zhang , Xin Luo , Li Li , Dong Liu This is my paper
classification eess.IVcs.CV
keywords extremeimagecompressionone-stepdiffusionSD-Turbolatentmodelrate-distortion-perceptiontradeoffgenerativecodecentropycodingultra-lowbitrate
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to show that extreme image compression can be made practical by sending a noisy latent representation over the bitstream and letting a single denoising step of a pre-trained text-to-image diffusion model invert that noise into a realistic image. Prior diffusion codecs needed dozens of denoising steps at the decoder, making them slow, and often drifted from the original image's content. StableCodec instead transmits the noisy latent of SD-Turbo at bitrates as low as 0.005 bits per pixel, and augments the pipeline with an auxiliary encoder and decoder to preserve structure and fidelity. The authors report state-of-the-art FID, KID, and DISTS scores on CLIC 2020 and DIV2K at these extreme bitrates, with decoding times comparable to mainstream neural codecs. If the claim holds, real-time generative compression at extreme rates becomes feasible rather than a research curiosity.

What carries the argument

The load-bearing mechanism is the one-step denoising inversion of Eq. (3): the decoded latent $l_T$ is fed to SD-Turbo's denoiser $\epsilon_{SD}$ at timestep $T$, and the formula $l_0 = (l_T - \sqrt{1-\bar\alpha_T}\,\epsilon_{SD}(l_T,T))/\sqrt{\bar\alpha_T}$ produces the clean latent in a single step, exploiting the assumption that $l_T$ has the noise level of the DDPM forward process. Around this sit the Deep Compression Latent Codec (extreme analysis/synthesis transforms with a 4-step quadtree autoregressive entropy model and latent residual prediction) that squeezes the latent to ultra-low bitrate, and the Dual-Branch Coding Structure (frozen ELIC-based auxiliary encoder $E_{Aux}$ for entropy-aware semantics plus auxiliary decoder $D_{Aux}$ for structure apportionment) that keeps fidelity from collapsing at those rates. The LoRA-adapted SD-Turbo supplies the generative prior while the VAE decoder stays frozen to preserve it.

What would settle it

Measure the decoded latent $l_T$ at 0.005 bpp against the Gaussian noise assumed by the DDPM schedule at step $T$ (for example, compare channel-wise variances of $l_T$ with the assumed noise level $1-\bar\alpha_T$), then rescale $l_T$ so it exactly matches the assumed noise level and rerun the one-step inversion. If FID/KID barely change after the correction, the schedule assumption is doing no load-bearing work; if they improve substantially, the assumption is violated at operating bitrates and the reported regime is narrower than claimed.

Watch

Extended reading notes

Core claim

The central claim is that a noisy latent $l_T$ reconstructed from an ultra-low-bitrate bitstream can be treated as if it were the $T$-th step of a DDPM forward noising process, so that the single-step inversion $l_0 = (l_T - \sqrt{1-\bar\alpha_T}\,\epsilon_{SD}(l_T,T))/\sqrt{\bar\alpha_T}$ followed by the frozen VAE decoder produces a high-realism, high-fidelity image. To make this work at 0.005 bpp, the paper introduces a Deep Compression Latent Codec that entropy-codes the latent at 64$\times$ spatial compression (hyperprior at 256$\times$), a frozen ELIC analysis transform as an auxiliary encoder for pixel-level semantic information, and an auxiliary decoder that routes basic structure directly from the bitstream so the denoiser can focus on generating realistic details. With end-to-end training that jointly optimizes bitrate, MSE/LPIPS/CLIP distortion, and an adversarial loss, the authors report that StableCodec outperforms existing codecs including prior diffusion-based ones on perceptual and structural similarity metrics while preserving fidelity and matching mainstream transform codecs in speed.

Load-bearing premise

The pipeline rests on the assumption that a latent quantized and compressed to about 0.005 bits per pixel still looks enough like ordinary Gaussian diffusion noise that one denoising step can invert it cleanly; if the compression noise strays too far from that assumed noise, reconstruction quality collapses.

Editorial extensions

If this is right

  • Diffusion-based codecs no longer need multi-step denoising, removing the main latency obstacle to real-time generative compression.
  • Bitrates around 0.005 bpp become usable for photo-realistic reconstruction of natural images, not just thumbnails.
  • The recipe of freezing the VAE decoder while adapting the encoder and denoiser with LoRA preserves a text-to-image model's priors while repurposing it for compression.
  • The auxiliary decoder demonstrates a concrete way to split decoding between structural information from the bitstream and generative detail from the denoiser.
  • At extreme bitrates, FID, KID, and DISTS are the metrics that track visual quality, while PSNR and MS-SSIM mislead.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 96-bit-per-image color-fix side information from the supplementary suggests a cheap, general remedy for color drift in tiled diffusion decoding that other latent-diffusion codecs could borrow.
  • The approach makes a testable prediction: quality should drop sharply when the decoded latent's noise distribution departs from the DDPM schedule, so the practical floor of the method is set by how far quantization noise can be pushed before that departure becomes large.
  • A natural extension the paper does not explore is video, where one-step decoding at transform-codec speeds could make low-bitrate generative video coding practical.
  • Using a frozen ELIC analysis transform as the auxiliary encoder hints that rate-distortion-oriented codecs and generative priors can be composed modularly instead of trained together from scratch.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. StableCodec is a learned image codec for extreme ultra-low bitrates (down to 0.005 bpp). It compresses a noisy latent representation of the input image using a deep transform codec in the VAE latent space of SD-Turbo, then reconstructs the image with a single denoising step. A dual-branch auxiliary encoder/decoder is introduced to improve reconstruction fidelity, and training uses a two-stage objective combining rate, MSE, LPIPS, CLIP distance, and a GAN loss. The paper reports state-of-the-art FID, KID, and DISTS curves on CLIC 2020 and DIV2K, along with a user study and runtime analysis showing inference speeds comparable to mainstream neural codecs.

Significance. If the reported margins are robust and the comparison is fair, StableCodec is a significant advance: it demonstrates that one-step diffusion can be used for extreme image compression with realistic, high-fidelity reconstructions, while overcoming the multi-step decoding bottleneck of prior diffusion codecs. The paper is thorough in its ablations (Tables 3-5), user study (Table 7), and runtime analysis (Tables 2 and 6), and the authors state that source code will be released. The method does not fit free parameters to the test sets, and the GAN/CLIP losses are training objectives rather than circular evaluation. The main risk to the central claim is the undisclosed color-correction side channel described in the supplementary, which could account for a nontrivial part of the perceptual metric gains.

major comments (2)
  1. [Supplementary A (Eqs. 9-11) and Section 4.2 (Fig. 6)] The main results appear to include a per-image global color-correction post-process (16-bit quantized mean/variance adaptive instance normalization) that is described only in the supplementary. No compared baseline receives the same 96-bit side information, and the main text does not disclose that StableCodec results in Fig. 6 and Fig. 7 use this step. Because DISTS is sensitive to global color statistics and FID/KID can also be improved by aligning the reconstruction's color distribution with the original, the reported SOTA margins may reflect this side channel rather than the proposed codec. The paper must either apply the identical correction to all baselines or provide an ablation quantifying the contribution of the color fix; otherwise the central claim is not supported.
  2. [Section 3.2, Eq. (3)] The one-step denoising formula inverts the DDPM forward process under the assumption that the transmitted latent l_T equals l0 plus Gaussian noise at the T-th schedule. At extreme bitrates (0.005 bpp) the quantization/compression error is non-Gaussian and likely large. The paper does not measure the distribution mismatch between the actual decoded l_T and the assumed noise level, nor does it analyze how the method degrades as this assumption breaks (e.g., sensitivity to the choice of T). Since the end-to-end training may compensate for the mismatch, this is not necessarily a fatal flaw, but the paper should provide such an analysis or justify the assumption to support the claimed mechanism.
minor comments (5)
  1. [Section 4.2.1 and Fig. 6] Please provide a table with numerical FID/KID/DISTS values (and ideally bootstrap confidence intervals) so the 'significant margin' claim can be quantified; the curves alone do not support precise comparisons.
  2. [Section 4.1] The sentence 'we neglect the results on Kodak as it is too small for calculating' should specify that this applies to FID/KID only, since DISTS and LPIPS are still reported for Kodak.
  3. [Section 3.2] There is a typo: '4-step antoregressive' should be 'autoregressive'; also Section 4.1 has 'repectively' instead of 'respectively'.
  4. [Table 5] The positive BD-rate values (24.41%, 13.29%) for PSNR and MS-SSIM in the rows with Ladv are easy to misread as improvements; add a '+' sign or a note that positive values denote degradation.
  5. [Supplementary Fig. 10] The notation for variance σ is overloaded with the entropy-model parameters; consider using a different symbol to avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the one-step denoising codec is trained end-to-end on a held-out training split and evaluated on external benchmarks; the per-image color-correction side-channel is a comparison-fairness caveat, not a circular derivation step.

full rationale

The central chain—compress a noisy latent with ga/gs, transmit and decode it, invert with the pretrained SD-Turbo denoiser in Eq. (3), and combine with the auxiliary branch in Eq. (4)—does not reduce to its own inputs. Eq. (3) is the standard DDPM single-step inversion l0 = (lT - sqrt(1-αbar_T) * epsilon_SD(lT, T)) / sqrt(αbar_T); it algebraically rearranges the forward noising model and uses the frozen pretrained denoiser, so it is not defined in terms of the target reconstruction. The paper trains on DF2K and CLIC 2020 Professional Training and evaluates on CLIC 2020 Test, DIV2K Val, and Kodak, so no test-set parameter is fitted. Optimizing MSE, LPIPS, CLIP, and GAN losses during training and then reporting FID, KID, DISTS, and LPIPS on held-out sets is a training-objective overlap, not a fitted-input-called-prediction pattern: the reported metrics are computed from reconstructions of test images that were not used in training, and no metric value is directly enforced as a trainable parameter on those images. The ablations in Tables 3-5 independently support the contribution of EAux, DAux, LoRA ranks, and the training strategy. The supplementary color fix (Supp. Eqs. 9-11) transmits the original image's per-channel mean and variance (96 bits per image) and applies adaptive instance normalization; this is a disclosed side-channel that can affect color-sensitive full-reference metrics and should ideally be applied or ablated across baselines, but it is not an equation in the claimed derivation chain and does not make Eq. (3) equivalent to Eq. (9). Similarly, substituting PerCo (SD) for PerCo and retraining or finetuning some baselines is a reported comparison-fairness limitation, not a circular step. No load-bearing self-citation, imported uniqueness theorem, or ansatz-smuggled-via-citation chain was found; the derivation is self-contained relative to the external SD-Turbo and ELIC priors it invokes.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim rests on three unverified assumptions: the one-step denoising validity, the Gaussian-noise treatment of the compressed latent, and the transferability of the frozen ELIC features. These are not derived or benchmarked independently; they are evaluated only through the final image-quality metrics.

free parameters (4)
  • One-step denoising timestep T = not reported
    Eq. (3) uses the noise schedule alpha_bar_T to convert l_T into l0, but the paper never states the value of T; the noise level of the transmitted latent depends on this hand-chosen hyperparameter.
  • Loss weights d1, d2, d3, beta = 2, 1, 0.1, 0.1
    Set in Section 4.1 by hand; these balance MSE, LPIPS, CLIP and GAN losses, and directly shape the fidelity-perception operating point claimed in the paper.
  • Bitrate Lagrange multipliers lambda_base and lambda_target = 0.5; {2,3,4,6,8,12,16,24,32}
    Two-stage implicit bitrate pruning uses lambda_base for warm-up and lambda_target for fine-tuning; the choice of this set determines the rate points on the reported RD curves.
  • LoRA ranks = 16 (ESD) and 32 (epsilonSD)
    Chosen by ablation (Table 4); larger ranks degrade performance, so this hand-selected value affects the trade-off between adapting the prior and preserving it.
assumptions (3)
  • domain assumption SD-Turbo can denoise a severely compressed latent in one step and preserve the original image content.
    The entire method depends on the pre-trained SD-Turbo one-step denoiser (Section 3.1); the paper cites [72] for this capability but does not analyze when it fails.
  • ad hoc to paper The decoded latent l_T behaves like a standard noised latent from the DDPM forward process.
    Eq. (3) inverts the noise using the DDPM schedule alpha_bar_T; the compression/quantization noise is not Gaussian, so this is an assumption specific to this paper that is not directly tested.
  • domain assumption The frozen analysis transform of a pre-trained ELIC model provides entropy-aware features useful at ultra-low bitrates.
    EAux is a frozen ELIC encoder (Section 3.3.1, Fig. 3); the paper assumes its features are useful at 0.005 bpp even though ELIC was trained for much higher rates.

how reviews work

0 comments
Cite this review

Pith. "Pith review of StableCodec: Taming One-Step Diffusion for Extreme Image Compression." pith.science (2026). https://pith.science/paper/TQ2QPTFU

@misc{pith2026250621977,
  author       = {Pith},
  title        = {Pith review of: StableCodec: Taming One-Step Diffusion for Extreme Image Compression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TQ2QPTFU}},
  note         = {Machine review of arXiv:2506.21977}
}
read the original abstract

Diffusion-based image compression has shown remarkable potential for achieving ultra-low bitrate coding (less than 0.05 bits per pixel) with high realism, by leveraging the generative priors of large pre-trained text-to-image diffusion models. However, current approaches require a large number of denoising steps at the decoder to generate realistic results under extreme bitrate constraints, limiting their application in real-time compression scenarios. Additionally, these methods often sacrifice reconstruction fidelity, as diffusion models typically fail to guarantee pixel-level consistency. To address these challenges, we introduce StableCodec, which enables one-step diffusion for high-fidelity and high-realism extreme image compression with improved coding efficiency. To achieve ultra-low bitrates, we first develop an efficient Deep Compression Latent Codec to transmit a noisy latent representation for a single-step denoising process. We then propose a Dual-Branch Coding Structure, consisting of a pair of auxiliary encoder and decoder, to enhance reconstruction fidelity. Furthermore, we adopt end-to-end optimization with joint bitrate and pixel-level constraints. Extensive experiments on the CLIC 2020, DIV2K, and Kodak dataset demonstrate that StableCodec outperforms existing methods in terms of FID, KID and DISTS by a significant margin, even at bitrates as low as 0.005 bits per pixel, while maintaining strong fidelity. Additionally, StableCodec achieves inference speeds comparable to mainstream transform coding schemes. All source code are available at https://github.com/LuizScarlet/StableCodec.

Figures

Figures reproduced from arXiv: 2506.21977 by the authors.

Figure 1
Figure 1. Visual examples and comparisons when compressing a 4K-resolution image [39] at ultra-low bitrates. The proposed Stable￾Codec produces more realistic and consistent details with fewer bits. In contrast, VVC [11], ELIC [22] and MS-ILLM [46] reconstructions are blurry, while PerCo [12] and DiffEIC [40] generate inconsistent details against the original images. Best viewed on screen for details. Abstract Diffusion-based… view at source ↗
Figure 2
Figure 2. (Top) Illustration of our motivation. One-step dif￾fusion can produce perceptually consistent results given severely corrupted images and a general prompt. (Bottom) Challenges in StableCodec. How to compress a noisy latent for one-step diffu￾sion using ultra-low bitrates, and how to improve fidelity. bitrate as low as 0.003 bits per pixel (bpp) using a pre￾trained latent diffusion model (LDM), highlighting the po￾te… view at source ↗
Figure 3
Figure 3. The framework of StableCodec. We incorporate the proposed Deep Compression Latent Codec to transmit a noisy latent lT for one-step denoising, where 64× denotes a spatial compression ratio of 64. To adjust the latent resolution, we deploy DownSample block and Conv3×3 as adapters after the VAE encoder ESD and auxiliary encoder EAux, respectively. We use a general prompt in both training and inference. The auxiliary de… view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Top-energy channels learned from different encoders. EAux embeds more pixel-level semantic information into codec. and provides rich, entropy-aware semantic information of the input images. To combine the latents from different en￾coders, we introduce trainable adapter…
Figure 6
Figure 6. Figure 6: Rate-distortion and rate-perception curve comparisons of different methods on the CLIC 2020, DIV2K and Kodak dataset. x and xˆ. We use DINOv2 [47] with registers [15] as the discriminator backbone [35]. To ensure stable training, we incorporate the GAN only in the seco…
Figure 7
Figure 7. Figure 7: Qualitative comparisons of different methods on Kodak. Best viewed on screen for details. the results on Kodak as it is too small for calculating. Note that pixel-level distortion metrics like LPIPS, PSNR and MS-SSIM have strong limitations when evaluating image compre…
Figure 8
Figure 8. Figure 8: Visual examples of color fix from CLIC 2020 [60]. 16- bit color fix brings clear refinement with negligible bits increase. built with a hyperprior module and an autoregressive con￾text model, where we first obtain and transmit a hyperprior Φhyper from y using the hyper…
Figure 9
Figure 9. Figure 9: (Left) Illustration of the entropy model. We build our entropy model on the basis of [45], which consists of a pair of hyper transforms, ha and hs, and a context model to perform entropy estimation for yˆ in an autoregressive manner. (Right) Illustration of the 4-step …
Figure 10
Figure 10. Figure 10: Module structures and network details. Method Encoding Time (s) Decoding Time (s) ESD EAux ga EE ED gs DAux ϵSD DSD StableCodec (Ours) 0.108 0.014 0.005 0.029 0.041 0.004 0.004 0.112 0.161 ELIC [22] - - 0.015 0.138 0.230 0.016 - - - [PITH_FULL_IMAGE:figures/full_fig_…
Figure 11
Figure 11. Figure 11: Additional rate-distortion curves on CLIC 2020 [60] and DIV2K [2] in terms of PSNR and MS-SSIM. Method HiFiC MS-ILLM Text+Sketch PerCo DiffEIC EGIC TACO StableCodec (Ours) Bitrate (bpp) 0.0268 0.0262 0.0274 0.0321 0.0375 0.0247 0.0258 0.0250 Top-1 Votes 20 26 11 24 43…
Figure 12
Figure 12. Figure 12: Visual examples and comparisons on 2K-resolution images from DIV2K. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]
Figure 13
Figure 13. Figure 13: Visual examples and comparisons on 2K-resolution images from CLIC 2020. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]
Figure 14
Figure 14. Figure 14: Visual examples and comparisons on 4K-resolution images from USTC-TD [39]. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_14.png]
Figure 15
Figure 15. Figure 15: Visual examples and comparisons on 4K-resolution images from USTC-TD [39]. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_15.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Next-Frame Decoding for Ultra-Low-Bitrate Image Compression with Video Diffusion Priors

    cs.CV 2026-03 conditional novelty 6.5 of 10

    Ultra-low-bitrate image decoding is cast as one-step next-frame prediction from a compact anchor using adapted video diffusion priors, yielding large perceptual bitrate savings versus DiffC.

Reference graph

Works this paper leans on

74 extracted references · 46 canonical work pages · cited by 1 Pith paper

  1. [1]

    https://github

    Tiled diffusion & vae extension. https://github. com/pkuliyi2015/multidiffusion-upscaler- for-automatic1111, 2023. Accessed: 2024-08-27. 12

    work page 2023

  2. [2]

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

    Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Pro- ceedings of the IEEE conference on computer vision and pat- tern recognition workshops, pages 126–135, 2017. 3, 6, 14

    work page 2017

  3. [3]

    Generative adversar- ial networks for extreme learned image compression

    Eirikur Agustsson, Michael Tschannen, Fabian Mentzer, Radu Timofte, and Luc Van Gool. Generative adversar- ial networks for extreme learned image compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 221–231, 2019. 2, 3

    work page 2019

  4. [4]

    Multi-realism image compression with a conditional generator

    Eirikur Agustsson, David Minnen, George Toderici, and Fabian Mentzer. Multi-realism image compression with a conditional generator. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 22324–22333, 2023. 2, 3

    work page 2023

  5. [5]

    End-to-end optimized image compression

    Johannes Ball ´e, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. arXiv preprint arXiv:1611.01704, 2016. 2, 3

    arXiv 2016

  6. [6]

    Variational image compression with a scale hyperprior

    Johannes Ball ´e, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436 ,

    arXiv

  7. [7]

    Demystifying mmd gans

    Mikołaj Bi ´nkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 3, 6

    arXiv 2018

  8. [8]

    Calculation of average psnr differences be- tween rd-curves

    G Bjontegaard. Calculation of average psnr differences be- tween rd-curves. ITU-T SG16 Q, 6, 2001. 8

    work page 2001

Show all 74 references
  1. [9]

    The perception-distortion tradeoff

    Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228–6237, 2018. 2, 3

    2018

  2. [10]

    Rethinking lossy com- pression: The rate-distortion-perception tradeoff

    Yochai Blau and Tomer Michaeli. Rethinking lossy com- pression: The rate-distortion-perception tradeoff. In Inter- national Conference on Machine Learning , pages 675–685. PMLR, 2019. 2, 3

    2019

  3. [11]

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

    Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Tech- nology, 31(10):3736–3764, 2021. 1, 2, 3, 7

    2021

  4. [12]

    Towards image compression with per- fect realism at ultra-low bitrates

    Marlene Careil, Matthew J Muckley, Jakob Verbeek, and St´ephane Lathuili`ere. Towards image compression with per- fect realism at ultra-low bitrates. InThe Twelfth International Conference on Learning Representations, 2023. 1, 2, 3, 7, 8, 12, 14

    2023

  5. [13]

    Learned image compression with discretized gaussian mixture likelihoods and attention modules

    Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939–7948, 2020. 2, 3

    2020

  6. [14]

    Perception pri- oritized training of diffusion models

    Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception pri- oritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472–11481, 2022. 12

    2022

  7. [15]

    Vision transformers need registers

    Timoth ´ee Darcet, Maxime Oquab, Julien Mairal, and Pi- otr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 6

    2023 arXiv

  8. [16]

    Image quality assessment: Unifying structure and texture similarity

    Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and ma- chine intelligence, 44(5):2567–2581, 2020. 3, 6, 7, 14

    2020

  9. [17]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873–12883, 2021. 3

    2021

  10. [18]

    Kodak lossless true color image suite (pho- tocd pcd0992)

    Rich Franzen. Kodak lossless true color image suite (pho- tocd pcd0992). http://r0k.us/graphics/kodak/,

  11. [19]

    Generative adversarial networks

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Commu- nications of the ACM, 63(11):139–144, 2020. 3

    2020

  12. [20]

    Causal contextual prediction for learned image com- pression

    Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Causal contextual prediction for learned image com- pression. IEEE Transactions on Circuits and Systems for Video Technology, 32(4):2329–2341, 2021. 2, 3, 4

    2021

  13. [21]

    Checkerboard context model for effi- cient learned image compression

    Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for effi- cient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771–14780, 2021. 2, 3

    2021

  14. [22]

    Elic: Efficient learned image compres- sion with unevenly grouped space-channel contextual adap- tive coding

    Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compres- sion with unevenly grouped space-channel contextual adap- tive coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 57...

    2022

  15. [23]

    Po-elic: Perception-oriented efficient learned image coding

    Dailan He, Ziming Yang, Hongjiu Yu, Tongda Xu, Jixiang Luo, Yuan Chen, Chenjian Gao, Xinjie Shi, Hongwei Qin, and Yan Wang. Po-elic: Perception-oriented efficient learned image coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages...

    2022

  16. [24]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium. Advances in neural information processing systems , 30, 2017. 3, 6

    2017

  17. [25]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020. 3, 4

    2020

  18. [26]

    High- fidelity image compression with score-based generative models

    Emiel Hoogeboom, Eirikur Agustsson, Fabian Mentzer, Luca Versari, George Toderici, and Lucas Theis. High- fidelity image compression with score-based generative models. arXiv preprint arXiv:2305.18231, 2023. 2, 3

    2023 arXiv

  19. [27]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3

    2021 arXiv

  20. [28]

    Generative latent coding for ultra-low bitrate image com- pression

    Zhaoyang Jia, Jiahao Li, Bin Li, Houqiang Li, and Yan Lu. Generative latent coding for ultra-low bitrate image com- pression. In Proceedings of the IEEE/CVF Conference on 9 Computer Vision and Pattern Recognition , pages 26088– 26098, 2024. 3, 4, 6, 7, 14

    2024

  21. [29]

    Mlic++: Linear complex- ity multi-reference entropy modeling for learned image com- pression

    Wei Jiang and Ronggang Wang. Mlic++: Linear complex- ity multi-reference entropy modeling for learned image com- pression. In ICML 2023 Workshop Neural Compression: From Information Theory to Applications, 2023. 7, 8

    2023

  22. [30]

    Mlic: Multi-reference entropy model for learned image compression

    Wei Jiang, Jiayu Yang, Yongqi Zhai, Peirong Ning, Feng Gao, and Ronggang Wang. Mlic: Multi-reference entropy model for learned image compression. In Proceedings of the 31st ACM International Conference on Multimedia , pages 7618–7627, 2023. 2, 3, 4

    2023

  23. [31]

    Mixture of diffusers for scene composition and high resolution image generation

    ´Alvaro Barbero Jim ´enez. Mixture of diffusers for scene composition and high resolution image generation. arXiv preprint arXiv:2302.02412, 2023. 8, 12

    2023 arXiv

  24. [32]

    Auto-encoding variational bayes

    Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3

    2013 arXiv

  25. [33]

    Egic: enhanced low-bit-rate generative image compression guided by semantic segmentation

    Nikolai K ¨orber, Eduard Kromer, Andreas Siebert, Sascha Hauke, Daniel Mueller-Gritschneder, and Bj ¨orn Schuller. Egic: enhanced low-bit-rate generative image compression guided by semantic segmentation. In European Conference on Computer Vision, pages 202–220. Springer, 2024...

    2024

  26. [34]

    Perco (sd): Open perceptual compression

    Nikolai K ¨orber, Eduard Kromer, Andreas Siebert, Sascha Hauke, Daniel Mueller-Gritschneder, and Bj ¨orn Schuller. Perco (sd): Open perceptual compression. arXiv preprint arXiv:2409.20255, 2024. 7, 14

    2024 arXiv

  27. [35]

    Ensembling off-the-shelf models for gan training

    Nupur Kumari, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Ensembling off-the-shelf models for gan training. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 10651–10662, 2022. 6

    2022

  28. [36]

    Neural image compres- sion with text-guided encoding for both pixel-level and per- ceptual fidelity

    Hagyeong Lee, Minkyu Kim, Jun-Hyuk Kim, Seungeon Kim, Dokwan Oh, and Jaeho Lee. Neural image compres- sion with text-guided encoding for both pixel-level and per- ceptual fidelity. arXiv preprint arXiv:2403.02944, 2024. 5, 7, 14

    2024 arXiv

  29. [37]

    Text+ sketch: Image compression at ultra low rates

    Eric Lei, Yi ˘git Berkay Uslu, Hamed Hassani, and Shirin Saeedi Bidokhti. Text+ sketch: Image compression at ultra low rates. arXiv preprint arXiv:2307.01944, 2023. 2, 3, 7, 8, 12, 14

    2023 arXiv

  30. [38]

    Neural video compression with diverse contexts

    Jiahao Li, Bin Li, and Yan Lu. Neural video compression with diverse contexts. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 22616–22626, 2023. 4, 12, 13

    2023

  31. [39]

    Ustc-td: A test dataset and bench- mark for image and video coding in 2020s

    Zhuoyuan Li, Junqi Liao, Chuanbo Tang, Haotian Zhang, Yuqi Li, Yifan Bian, Xihua Sheng, Xinmin Feng, Yao Li, Changsheng Gao, et al. Ustc-td: A test dataset and bench- mark for image and video coding in 2020s. arXiv preprint arXiv:2409.08481, 2024. 1, 14, 17, 18

    2024 arXiv

  32. [40]

    Towards extreme image compression with latent feature guidance and diffusion prior

    Zhiyuan Li, Yanhui Zhou, Hao Wei, Chenyang Ge, and Jing- wen Jiang. Towards extreme image compression with latent feature guidance and diffusion prior. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1, 2, 3, 4, 6, 7, 8, 12, 14

    2024

  33. [41]

    Enhanced deep residual networks for single image super-resolution

    Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE confer- ence on computer vision and pattern recognition workshops, pages 136–144, 2017. 6

    2017

  34. [42]

    Learned image compression with mixed transformer-cnn architectures

    Jinming Liu, Heming Sun, and Jiro Katto. Learned image compression with mixed transformer-cnn architectures. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 14388–14397, 2023. 2, 3, 4

    2023

  35. [43]

    High-fidelity generative image compres- sion

    Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compres- sion. Advances in Neural Information Processing Systems , 33:11913–11924, 2020. 2, 3, 7, 8, 14

    2020

  36. [44]

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

    David Minnen and Saurabh Singh. Channel-wise autoregres- sive entropy models for learned image compression. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3339–3343. IEEE, 2020. 2, 3, 4, 5, 12, 13

    2020

  37. [45]

    Joint autoregressive and hierarchical priors for learned im- age compression

    David Minnen, Johannes Ball ´e, and George D Toderici. Joint autoregressive and hierarchical priors for learned im- age compression. Advances in neural information processing systems, 31, 2018. 2, 3, 12, 13

    2018

  38. [46]

    Improving statistical fi- delity for neural image compression with implicit local like- lihood models

    Matthew J Muckley, Alaaeldin El-Nouby, Karen Ullrich, Herv´e J ´egou, and Jakob Verbeek. Improving statistical fi- delity for neural image compression with implicit local like- lihood models. In International Conference on Machine Learning, pages 25426–25443. PMLR, 2023. 1, 2,...

    2023

  39. [47]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6

    2023 arXiv

  40. [48]

    One-step image translation with text-to-image models

    Gaurav Parmar, Taesung Park, Srinivasa Narasimhan, and Jun-Yan Zhu. One-step image translation with text-to-image models. arXiv preprint arXiv:2403.12036, 2024. 3

    2024 arXiv

  41. [49]

    Entroformer: A transformer-based entropy model for learned image compression

    Yichen Qian, Ming Lin, Xiuyu Sun, Zhiyu Tan, and Rong Jin. Entroformer: A transformer-based entropy model for learned image compression. arXiv preprint arXiv:2202.05492, 2022. 3, 4

    2022 arXiv

  42. [50]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  43. [51]

    Lossy image compression with foundation diffusion models

    Lucas Relic, Roberto Azevedo, Markus Gross, and Christo- pher Schroers. Lossy image compression with foundation diffusion models. In European Conference on Computer Vi- sion, pages 303–319. Springer, 2024. 2, 3

    2024

  44. [52]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2, 3, 4

    2022

  45. [53]

    Progressive distillation for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3

    2022 arXiv

  46. [54]

    Adversarial diffusion distillation

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision , pages 87–103. Springer,

  47. [55]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International confer- ence on machine learning, pages 2256–2265. PMLR, 2015. 3

    2015

  48. [56]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3

    2010 arXiv

  49. [57]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 3

    2023 arXiv

  50. [58]

    Overview of the high efficiency video coding (hevc) standard

    Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology, 22(12):1649–1668, 2012. 3

    2012

  51. [59]

    Lossy compression with gaussian diffusion

    Lucas Theis, Tim Salimans, Matthew D Hoffman, and Fabian Mentzer. Lossy compression with gaussian diffusion. arXiv preprint arXiv:2206.08889, 2022. 2, 3

    2022 arXiv

  52. [60]

    Clic 2020: Challenge on learned image compression, 2020, 2020

    George Toderici, Lucas Theis, Nick Johnston, Eirikur Agustsson, Fabian Mentzer, Johannes Ball ´e, Wenzhe Shi, and Radu Timofte. Clic 2020: Challenge on learned image compression, 2020, 2020. 3, 6, 12, 14

    2020

  53. [61]

    Neural discrete representation learning

    Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information pro- cessing systems, 30, 2017. 3

    2017

  54. [62]

    The jpeg still picture compression stan- dard

    Gregory K Wallace. The jpeg still picture compression stan- dard. Communications of the ACM , 34(4):30–44, 1991. 2, 3

    1991

  55. [63]

    Exploiting diffusion prior for real-world image super-resolution

    Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1–21, 2024. 3, 8, 12

    2024

  56. [64]

    Mul- tiscale structural similarity for image quality assessment

    Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Mul- tiscale structural similarity for image quality assessment. In The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398–1402. Ieee, 2003. 6

    2003

  57. [65]

    Idempotence and perceptual image com- pression

    Tongda Xu, Ziran Zhu, Dailan He, Yanghao Li, Lina Guo, Yuanyuan Wang, Zhe Wang, Hongwei Qin, Yan Wang, Jingjing Liu, et al. Idempotence and perceptual image com- pression. arXiv preprint arXiv:2401.08920, 2024. 2, 3

    2024 arXiv

  58. [66]

    On perceptual lossy compression: The cost of percep- tual reconstruction and an optimal training framework

    Zeyu Yan, Fei Wen, Rendong Ying, Chao Ma, and Peilin Liu. On perceptual lossy compression: The cost of percep- tual reconstruction and an optimal training framework. InIn- ternational Conference on Machine Learning, pages 11682– 11692. PMLR, 2021. 2, 3

    2021

  59. [67]

    Optimally con- trollable perceptual lossy compression

    Zeyu Yan, Fei Wen, and Peilin Liu. Optimally con- trollable perceptual lossy compression. arXiv preprint arXiv:2206.10082, 2022. 2, 3

    2022 arXiv

  60. [68]

    Lossy image compression with conditional diffusion models

    Ruihan Yang and Stephan Mandt. Lossy image compression with conditional diffusion models. Advances in Neural In- formation Processing Systems, 36, 2024. 2, 3

    2024

  61. [69]

    Im- proved distribution matching distillation for fast image syn- thesis

    Tianwei Yin, Micha ¨el Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Im- proved distribution matching distillation for fast image syn- thesis. arXiv preprint arXiv:2405.14867, 2024. 3

    2024 arXiv

  62. [70]

    Mambaout: Do we really need mamba for vision? arXiv preprint arXiv:2405.07992,

    Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision? arXiv preprint arXiv:2405.07992,

    arXiv

  63. [71]

    Inceptionnext: When inception meets convnext

    Weihao Yu, Pan Zhou, Shuicheng Yan, and Xinchao Wang. Inceptionnext: When inception meets convnext. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5672–5683, 2024. 4, 12

    2024

  64. [72]

    Degradation-guided one-step im- age super-resolution with diffusion priors

    Aiping Zhang, Zongsheng Yue, Renjing Pei, Wenqi Ren, and Xiaochun Cao. Degradation-guided one-step im- age super-resolution with diffusion priors. arXiv preprint arXiv:2409.17058, 2024. 2, 3, 5

    2024 arXiv

  65. [73]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018. 5, 6 11 StableCodec:...

    2018

  66. [74]

    Sta- bleCodec - EAux

    for comparison, which only contains the analysis transform ga, the synthesis transform gs and the entropy model. and those of the entropy decoding, gs, auxiliary decoder DAux, one-step denoising Unet ϵSD and V AE decoderDSD during the decoding process. For comparison, we add t...

    2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.