Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Compressing an image into text, a highly compressed image, and learned pseudo-words and fusing them in a diffusion decoder keeps reconstructions semantically faithful below 0.05 bits per pixel.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A diffusion-based image codec guided by text, a highly compressed image, and CLIP-derived semantic pseudo-words improves semantic consistency at bitrates below 0.05 bpp.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection A competently engineered multimodal guidance framework for ultra-low-bitrate diffusion compression, but the abstract and body describe different methods, and the SPW bitrate accounting is load-bearing and under-specified. the 4 major comments →

arxiv 2512.06344 v2 pith:XYICI3RV submitted 2025-12-06 cs.CV

SDGIC: A Semantic Disambiguation-Guided Generative Image Compression Method for Ultra-Low Bitrates

Kaile Wang , Lijun He , Haisheng Fu , Haixia Bi , Fan Li This is my paper
classification cs.CV
keywords generative image compressionultra-low bitratesemantic consistencydiffusion modelmultimodal guidancesemantic pseudo-wordstask-aware compressionControlNet conditioning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

At bitrates below 0.05 bits per pixel, diffusion-based decoders given only sparse cues such as text or sketches tend to produce images that look natural but disagree with the source. This paper argues the fix is to send three complementary guidance streams: a short text caption for global meaning, a highly compressed image (HCI) for dense visual structure, and Semantic Pseudo-Words (SPWs) learned end-to-end toward the denoising objective for fine-grained, task-relevant semantics. A dual-path conditioned diffusion decoder injects these streams, using cross-attention for the semantic cues and ControlNet-style additive residuals for the HCI. The paper reports that this reduces semantic deviation at ultra-low bitrates—for example, 10.59% lower DISTS on DIV2K and 28.65% lower AFINE than a strong baseline on CLIC2020—while keeping perceptual quality and pixel fidelity competitive. A sympathetic reader would care because this offers a concrete recipe for making generative compression reliable enough for bandwidth-starved applications like 6G semantic communication.

Core claim

The central claim is that semantic inconsistency at ultra-low bitrate is caused by incomplete guidance, which leaves the diffusion decoder too much freedom to hallucinate confident but wrong content. To close the gap, the paper compresses the source image into three compact, complementary signals: a text caption (global semantics, ~20 words), an HCI generated by a standard learned codec (low-level structure and color), and Semantic Pseudo-Words produced by a Task-Aware Semantic Compression Module. The SPWs are a single 1024-dimensional vector in text-embedding space, trained end-to-end so that the multi-head self-attention of a small network distills exactly the semantics that reduce the den

What carries the argument

The load-bearing objects are the Task-Aware Semantic Compression Module (TASCM) and the Multimodal-Guided Diffusion Decoder (MGDD). TASCM turns a frozen visual-language encoder's CLS token into Semantic Pseudo-Words (SPWs)—a 1×1024 embedding aligned to the text-embedding space via a small MLP—by letting a trainable two-layer multi-head self-attention network select task-relevant semantics under the downstream denoising loss. MGDD is a dual-path conditioner: one path fuses the text caption and SPWs (SPWs replace a placeholder token in the caption) and injects them via cross-attention; the other path uses a trainable copy of the U-Net (ControlNet-style) to add multi-scale residuals from the HC

Load-bearing premise

The bitrate accounting assumes the semantic pseudo-word (one 1024-dimension vector), the ~20-word text, and the HCI together fit below 0.05 bpp, but the paper never specifies how the floating-point SPW is compressed to its reported ~0.0002 bpp; if near-lossless compression of the SPW is impossible at that rate, the ultra-low-bitrate claim gives way.

What would settle it

Measure the true transmitted size of the SPW at a 512×512 resolution. A single 1×1024 vector in half-precision floats costs 16,384 bits, which is 0.0625 bpp—already above the claimed 0.05 bpp ultra-low-bitrate range; at full precision it is double that. If the method's reported semantic-consistency gains require transmitting the SPW in a form that exceeds the budget, the central ultra-low-bitrate claim collapses. A direct test would compare DISTS/AFINE with SPWs sent losslessly vs. quantized to a realistic bitrate.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Each guidance stream plays a distinct role: text anchors global semantics, HCI anchors low-level structure and color, and SPWs supply task-relevant fine-grained constraints absent from the other two.
  • Optimizing SPWs directly against the denoising objective beats proxy-based semantic features, so end-to-end task-aware training is the load-bearing design choice.
  • The dual-path design (cross-attention for semantics, ControlNet residuals for spatial guidance) lets the frozen diffusion prior be steered without fine-tuning the whole network.
  • At ~0.03 bpp the method reports consistent gains over prior generative codecs (e.g., 10.59% lower DISTS on DIV2K; 28.65% lower AFINE vs PerCo on CLIC2020), which is the level of improvement needed for 6G semantic communication.
  • The framework stays training-efficient: only the task-aware module and the ControlNet are trained; the diffusion backbone is frozen.

Where Pith is reading between the lines

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

  • A natural testable extension is to replace the fixed-size pseudo-word with a resolution-adaptive, variable-length token sequence, which would preserve the reported gains for lower-resolution images where the fixed SPW cost currently dominates (the paper lists this as future work).
  • The task-aware pseudo-word idea is essentially textual inversion repurposed for compression—learning a per-image token that minimizes the reconstruction loss—which suggests it could transfer to other diffusion-conditioned tasks such as semantic editing or reference-based super-resolution.
  • Because the diffusion backbone and the HCI codec are frozen, the framework is modular: replacing the HCI codec with a better one (or improving the captioner) should directly improve the output without retraining the whole system—a property the paper does not explicitly claim.
  • The reliance on a large multimodal captioner means the method's semantic guarantees inherit the captioner's errors; an adversarial test would ask whether the method still outperforms baselines when captions are deliberately corrupted or ambiguous.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The full text proposes MTGC, a generative image compression framework for bpp<0.05. The encoder extracts three guidance streams: a concise text caption from an LMM, a highly compressed image (HCI) from ELIC, and one 1,024-dimensional Semantic Pseudo-Word (SPW) produced by TASCM. The decoder, MGDD, injects these streams into a frozen Stable Diffusion backbone via cross-attention for semantic conditions and ControlNet residuals for the HCI. A three-stage training procedure is described, and experiments on Tecnick, CLIC2020, and DIV2K claim consistent improvements in semantic consistency (DISTS, LPIPS, ClipScore, SeSS) and competitive perceptual/fidelity quality versus JPEG, VVC, ELIC, DCAE, HiFiC, VQIR, PICS, PerCo, and MISC. The arXiv abstract, however, describes a different framework named SDGIC with RSRTs and DPCD and reports a 23.4% AFINE reduction on CLIC2020, which does not match the body's MTGC/SPW/MGDD description and 10.59% DISTS claim.

Significance. The problem is timely and relevant: diffusion-based generative compression at ultra-low bitrate is a real bottleneck for semantic communication, and hallucination suppression is the central open issue. The proposed combination of a dense learned-codec image, a short caption, and a task-optimized pseudo-word embedding is a plausible and reasonably well-motivated approach. The experimental breadth is a strength: multiple datasets, several semantic/perceptual/fidelity metrics, ablations of each guidance component, denoising-step analysis, and decoding-latency comparisons. If the reported gains are reproducible, the framework is a useful system-level contribution. As submitted, however, the paper does not release code or checkpoints, reports no uncertainty estimates, and contains unresolved inconsistencies in method identity and bitrate accounting. The central idea is defensible and fixable, but the current manuscript cannot be accepted without addressing these points.

major comments (4)
  1. [Abstract vs §I, §III] The arXiv abstract and the full text describe different methods and different headline results. The abstract introduces "SDGIC" with "Reconstruction-Aware Semantic Residual Tokens (RSRTs)" and a "Dual-Path Conditioned Diffusion Decoder (DPCD)," and reports a 23.4% AFINE reduction on CLIC2020. The body proposes "MTGC" with "Semantic Pseudo-Words (SPWs)" and a "Multimodal-Guided Diffusion Decoder (MGDD)," and reports a 10.59% DISTS reduction on DIV2K. RSRTs and DPCD never appear in the body. This is not cosmetic: it is impossible to tell which system was actually evaluated. The title, abstract, method description, and headline numbers must be harmonized before the manuscript can be reviewed as a single coherent paper.
  2. [§III-B3, Fig. 2, Table II, §V-D] The ultra-low-bitrate claim rests on the SPW bitrate, but the paper never specifies how a continuous 1×1024 embedding is converted to a bitstream. Fig. 2 says SPWs are "losslessly compressed using Zstd" and §IV-A4 says half precision is used at inference, but no quantization or encoding details, and no measured compressed sizes, are given. A float16 1×1024 vector costs 16,384 bits, or 0.0156 bpp at 1024×1024; the reported 0.014 bpp is plausible only with a specific representation that is not described. At Tecnick's 768×768 resolution the same vector costs about 0.028 bpp in float16, or about 0.024 bpp if the same compression ratio as CLIC2020 is assumed, before adding the text caption and HCI. This can push the total above the advertised bpp<0.05 regime. The fixed-cost problem is acknowledged in §V-D, but Tecnick results are still presented as supporting the ultra-low-bitrate claim. Plea
  3. [§IV-B, Figs. 7–10] The central empirical claim is that MTGC "consistently improves" semantic consistency and achieves "remarkable gains" at bpp<0.05, but all reported values are single-run point estimates. No error bars, confidence intervals, or significance tests are provided for any metric, and code/checkpoints are not released. The headline percentages (e.g., 10.59% DISTS on DIV2K, roughly 28.65% AFINE on CLIC2020) therefore cannot be separated from training variance. Please report at least three seeds with standard deviations, or otherwise justify stability, and release evaluation code to make the comparisons reproducible.
  4. [§IV-A2 vs §III-B3] ClipScore is used as a semantic-consistency metric, yet the SPW extractor is built on the same frozen OpenCLIP ViT-H/14 visual encoder used to compute CLIP embeddings. Because the guidance is generated from the same representation family as one of the evaluation metrics, ClipScore gains partly measure self-consistency with the guidance encoder rather than independent reconstruction quality. Please either drop ClipScore from the semantic-consistency claims, add a CLIP-free semantic metric, or explicitly discuss why this alignment does not inflate the reported ClipScore. The other semantic metrics (DISTS, LPIPS, SeSS) should carry the main semantic claim.
minor comments (6)
  1. [§I, §III-C] Typos and terminology inconsistencies: "of of multimodal guidance" in §I, "ross-modal fusion" in §III-C, and "VCC" vs "VVC" in the related-work and Table I. Please proofread carefully.
  2. [Table III] The ablation compares MTGC-1 at 0.002 bpp with MTGC-2/3 at 0.039 bpp. The conclusion that text-only guidance is insufficient is confounded by the large bitrate difference. A text+SPW control at a matched rate, or an explicit discussion of this confound, would strengthen the ablation.
  3. [Fig. 11 and references] The qualitative figure caption lists VQIR as [21], but VQIR is reference [27] in the body. Please correct the citation and ensure all figure references match the reference list.
  4. [§IV-A3, MISC comparison] For MISC, results are taken from the original paper because code is unavailable. Please state explicitly whether the same test resolution, bitrate range, and caption protocol were used; otherwise the comparison is not controlled.
  5. [§IV-A1] Clarify whether the CC3M captions are used as-is or regenerated with Gemini 2.5 Pro. The caption protocol is a critical variable for all methods, so the source of the training captions should be explicit.
  6. [§VI] The conclusion claims the work "provides theoretical and technical foundations," but the paper is entirely empirical and contains no theoretical derivation. Please soften this claim to match the actual contribution.

Circularity Check

0 steps flagged

No significant circularity: the semantic-consistency claims rest on multiple independent metrics and no derivation step reduces to its own inputs.

full rationale

The paper is an empirical systems contribution, not a derivation from first principles. The three guidance streams (text, HCI, SPWs) are extracted from the source image and used to condition a frozen diffusion decoder; the claimed improvements are benchmark measurements. TASCM's SPWs are trained with the denoising objective (Eqs. 10-12); no evaluation metric is used as a training loss and no test-set parameter is fitted, so the fitted-input-called-prediction pattern does not apply. The only representation overlap is that SPWs are produced from an OpenCLIP ViT-H/14 encoder while one listed metric, ClipScore, is a CLIP image-feature cosine similarity. However, the paper does not specify that ClipScore uses the same CLIP model, and even if it did, the central semantic-consistency claim also rests on DISTS, LPIPS, SeSS, and AFINE, which are not derived from the SPW encoder. Thus no exact reduction Eq. X = Eq. Y can be exhibited. The Limitations section's admission that fixed-dimensionality SPWs raise bpp at lower resolution is a bitrate-accounting and external-validity risk, not a circularity. Self-citations ([42], [46]) are routine and not load-bearing. Verdict: no significant circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 1 invented entities

The central claim rests on standard diffusion/CLIP machinery and on the domain assumption that the three guidance streams—text, HCI, and SPWs—can be losslessly transmitted at negligible bitrate while providing complementary constraints. The SPW bitrate accounting is the least supported piece: no quantization or entropy coding details are given.

free parameters (6)
  • Number of SPWs L = 1
    Chosen by hand; fixed dimensionality 1×1024, affecting bitrate and capacity.
  • SemEnc self-attention layers T = 2
    Set to 2 in the implementation; a hyperparameter.
  • ELIC Lagrange multiplier λ_ELIC = 1, 2, 3 ×10⁻⁴
    Three values configure three operating bitrates; ablations use 2 or 3 ×10⁻⁴.
  • Denoising steps = 50
    Inference uses 50 steps; ablation shows step count affects metrics and complexity.
  • Text caption length = ≈20 words / ~0.002 bpp
    Chosen to balance semantic guidance against noise; ablation shows 20 words is optimal.
  • Learning rate schedule = 1e-4 → 5e-5
    Adam optimizer settings; part of training recipe needed for replication.
axioms (5)
  • domain assumption Pre-trained Stable Diffusion 2.1-base has sufficient generative priors to faithfully reconstruct images from multimodal guidance
    The whole decoder design relies on this prior; stated in Section III-C.
  • domain assumption CLIP visual embeddings encode task-relevant semantics that can be refined by a 2-layer self-attention network
    SemEnc premise in Section III-B-3.
  • domain assumption Concise text captions from an LMM provide macro-level semantic guidance without harmful semantic noise
    Prompt design in Section III-B-1; ablation Table IV.
  • standard math The diffusion forward process q(xt|x_{t-1}) in Eq. (1) is a valid model
    Standard DDPM formulation from Ho et al. [47].
  • domain assumption ELIC's H^{-1}(B(x)) reconstructs a useful HCI at the given λ
    HCI is generated by pre-trained ELIC; central to the second guidance path.
invented entities (1)
  • Semantic Pseudo-Words (SPWs) no independent evidence
    purpose: Fine-grained task-relevant semantic guidance injected via placeholder tokens in cross-attention
    Internal learned tokens; no external validation; effectiveness is only measured through the proposed system's downstream metrics. The abstract's RSRTs are absent from the manuscript body.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SDGIC: A Semantic Disambiguation-Guided Generative Image Compression Method for Ultra-Low Bitrates." pith.science (2026). https://pith.science/paper/XYICI3RV

@misc{pith2026251206344,
  author       = {Pith},
  title        = {Pith review of: SDGIC: A Semantic Disambiguation-Guided Generative Image Compression Method for Ultra-Low Bitrates},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XYICI3RV}},
  note         = {Machine review of arXiv:2512.06344}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Generative image compression has recently shown impressive perceptual quality, but often suffers from semantic inconsistency at ultra-low bitrates (bpp < 0.05), limiting its reliable deployment in bandwidth-constrained scenarios such as 6G semantic communications. This inconsistency stems from incomplete guidance information, which introduces semantic ambiguity into the generation process and may lead to natural-looking but source-inconsistent content. In this work, we propose a Semantic-Disambiguation-Guided Generative Image Compression (SDGIC) framework to constrain diffusion-based reconstruction at ultra-low bitrates. Specifically, SDGIC compresses the source image into three compact and complementary guidance streams: a concise text caption for global semantics, a highly compressed image (HCI) for dense visual evidence, and Reconstruction-Aware Semantic Residual Tokens (RSRTs) for reconstruction-relevant residual semantics that remain ambiguous under the text caption and HCI conditions. The RSRTs are directly optimized toward the downstream denoising objective, enabling them to provide source-specific semantic constraints for disambiguating diffusion-based reconstruction. To inject these three guidance streams into the generation process effectively, we design a Dual-Path Conditioned Diffusion Decoder (DPCD), which uses cross-attention for semantic conditions and ControlNet residuals for dense visual guidance. Extensive experiments demonstrate that SDGIC improves semantic consistency at ultra-low bitrates while maintaining favorable perceptual quality, with a 23.4% reduction in AFINE on the CLIC2020 dataset.

Figures

Figures reproduced from arXiv: 2512.06344 by Fan Li, Haisheng Fu, Haixia Bi, Kaile Wang, Lijun He.

Figure 1
Figure 1. Figure 1: Visual comparison of different methods at Ultra-low Bitrate. (b) ELIC [ [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall framework of the proposed MTGC. The encoder extracts three guidance modalities from the original image, including: a text caption, an HCI, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Detailed architecture of the TASCM, comprising a SemEnc and a [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Detailed architecture of MGDD, leveraging a pre-trained diffusion [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Detailed process of cross-modal fusion. A predefined placeholder is [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Quantitative comparison of semantic consistency metrics (CLIP score, LPIPS, DISTS, SeSS) evaluated on CLIC2020, DIV2K, and Tecnick datasets [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Quantitative comparison of the comprehensive metric AFINE eval [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Quantitative comparison of perceptual metrics (MANIQA, CLIPIQA) evaluated on CLIC2020, DIV2K, and Tecnick datasets with generative baselines [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Quantitative comparison of pixel-level fidelity metrics (PSNR, MS-SSIM) evaluated on CLIC2020, DIV2K, and Tecnick datasets with generative [PITH_FULL_IMAGE:figures/full_fig_p009_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative comparison with representative baselines (JPEG [ [PITH_FULL_IMAGE:figures/full_fig_p010_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative Comparison of Model Variants. MTGC-1 and MTGC-2 [PITH_FULL_IMAGE:figures/full_fig_p011_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Impact of Denoising Steps on Semantic Consistency. A quantitative ablation on CLIC2020 shows that increasing denoising steps consistently improves [PITH_FULL_IMAGE:figures/full_fig_p012_13.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. DecoyFace: Beyond Obfuscation via Controllable and Imperceptible Identity Misdirection for Privacy-Preserving Face Recognition

    cs.CV 2026-07 conditional novelty 6.0

    DecoyFace redirects inversion of split face-recognition features toward a plausible decoy identity while preserving authorized verification.

Reference graph

Works this paper leans on

64 extracted references · 8 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Semantic v2x communications for image transmission in 6g systems,

    J. M. Gimenez-Guzman, I. Leyva-Mayorga, and P. Popovski, “Semantic v2x communications for image transmission in 6g systems,”IEEE Network, vol. 38, no. 6, pp. 48–54, 2024

    2024

  2. [2]

    Quality-of-experience evaluation for digital twins in 6g network environments,

    Z. Zhang, Y . Zhou, L. Teng, W. Sun, C. Li, X. Min, X.-P. Zhang, and G. Zhai, “Quality-of-experience evaluation for digital twins in 6g network environments,”IEEE Transactions on Broadcasting, vol. 70, no. 3, pp. 995–1007, 2024

    2024

  3. [3]

    Learning-based data transmissions for future 6g enabled industrial iot: A data compression perspective,

    M. Zhang, H. Zhang, Y . Fang, and D. Yuan, “Learning-based data transmissions for future 6g enabled industrial iot: A data compression perspective,”IEEE Network, vol. 36, no. 5, pp. 180–187, 2022. 13

    2022

  4. [4]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE transactions on image processing, vol. 13, no. 4, pp. 600–612, 2004

    2004

  5. [5]

    Vari- ational image compression with a scale hyperprior,

    J. Ball ´e, D. Minnen, S. Singh, S. J. Hwang, and N. Johnston, “Vari- ational image compression with a scale hyperprior,”arXiv preprint arXiv:1802.01436, 2018

    Pith/arXiv arXiv 2018

  6. [6]

    End-to-end optimized image compression,

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

    Pith/arXiv arXiv 2016

  7. [7]

    Compressnet: Generative compression at extremely low bitrates,

    S. K. Raman, A. Ramesh, V . Naganoor, S. Dash, G. Kumaravelu, and H. Lee, “Compressnet: Generative compression at extremely low bitrates,” inProceedings of the IEEE/CVF Winter Conference on Appli- cations of Computer Vision, 2020, pp. 2325–2333

    2020

  8. [8]

    Generative compression,

    S. Santurkar, D. Budden, and N. Shavit, “Generative compression,” in 2018 Picture Coding Symposium (PCS). IEEE, 2018, pp. 258–262

    2018

  9. [9]

    Generative visual compression: A review,

    B. Chen, S. Yin, P. Chen, S. Wang, and Y . Ye, “Generative visual compression: A review,” in2024 IEEE International Conference on Image Processing (ICIP). IEEE, 2024, pp. 3709–3715

    2024

  10. [10]

    An optimal bit-rate allocation algorithm to improve transmission efficiency of images in deep space exploration,

    Y . Xu, J. Li, X. Lin, and F. Bai, “An optimal bit-rate allocation algorithm to improve transmission efficiency of images in deep space exploration,” China Communications, vol. 17, no. 7, pp. 94–100, 2020

    2020

  11. [11]

    Lossless compression techniques for low bandwidth net- works,

    S. K. Routray, A. Javali, K. Sharmila, W. Semunigus, M. Pappa, and A. D. Ghosh, “Lossless compression techniques for low bandwidth net- works,” in2020 3rd international conference on intelligent sustainable systems (ICISS). IEEE, 2020, pp. 823–828

    2020

  12. [12]

    The jpeg still picture compression standard,

    G. K. Wallace, “The jpeg still picture compression standard,”Commu- nications of the ACM, vol. 34, no. 4, pp. 30–44, 1991

    1991

  13. [13]

    Jpeg2000: Image compression fundamentals, standards and practice,

    D. S. Taubman, M. W. Marcellin, and M. Rabbani, “Jpeg2000: Image compression fundamentals, standards and practice,”Journal of Elec- tronic Imaging, vol. 11, no. 2, pp. 286–287, 2002

    2002

  14. [14]

    Bpg image format,

    F. Bellard, “Bpg image format,” 2015

    2015

  15. [15]

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

    B. Bross, Y .-K. Wang, Y . Ye, S. Liu, J. Chen, G. J. Sullivan, and J.- R. Ohm, “Overview of the versatile video coding (vvc) standard and its applications,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 31, no. 10, pp. 3736–3764, 2021

    2021

  16. [16]

    Elic: Efficient learned image compression with unevenly grouped space- channel contextual adaptive coding,

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

    2022

  17. [17]

    Learned image compression with gaussian- laplacian-logistic mixture model and concatenated residual modules,

    H. Fu, F. Liang, J. Lin, B. Li, M. Akbari, J. Liang, G. Zhang, D. Liu, C. Tu, and J. Han, “Learned image compression with gaussian- laplacian-logistic mixture model and concatenated residual modules,” IEEE Transactions on Image Processing, vol. 32, pp. 2063–2076, 2023

    2063

  18. [18]

    Frequency- aware transformer for learned image compression,

    H. Li, S. Li, W. Dai, C. Li, J. Zou, and H. Xiong, “Frequency- aware transformer for learned image compression,”arXiv preprint arXiv:2310.16387, 2023

    Pith/arXiv arXiv 2023

  19. [19]

    Preprocessing enhanced image compression for machine vision,

    G. Lu, X. Ge, T. Zhong, Q. Hu, and J. Geng, “Preprocessing enhanced image compression for machine vision,”IEEE transactions on circuits and systems for video technology, 2024

    2024

  20. [20]

    Learned image compression with hierarchical progressive context modeling,

    Y . Li, H. Zhang, L. Li, and D. Liu, “Learned image compression with hierarchical progressive context modeling,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 18 834–18 843

    2025

  21. [21]

    Learned image compression with dictionary-based entropy model,

    J. Lu, L. Zhang, X. Zhou, M. Li, W. Li, and S. Gu, “Learned image compression with dictionary-based entropy model,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 12 850–12 859

    2025

  22. [22]

    Text+ sketch: Image compression at ultra low rates,

    E. Lei, Y . B. Uslu, H. Hassani, and S. S. Bidokhti, “Text+ sketch: Image compression at ultra low rates,”arXiv preprint arXiv:2307.01944, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    High- fidelity generative image compression,

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

    2020

  24. [24]

    Cross modal compression: Towards human-comprehensible semantic compression,

    J. Li, C. Jia, X. Zhang, S. Ma, and W. Gao, “Cross modal compression: Towards human-comprehensible semantic compression,” inProceedings of the 29th ACM international conference on multimedia, 2021, pp. 4230–4238

    2021

  25. [25]

    Improving statistical fidelity for neural image compression with im- plicit local likelihood models,

    M. J. Muckley, A. El-Nouby, K. Ullrich, H. J ´egou, and J. Verbeek, “Improving statistical fidelity for neural image compression with im- plicit local likelihood models,” inInternational Conference on Machine Learning. PMLR, 2023, pp. 25 426–25 443

    2023

  26. [26]

    Extremely low bit-rate image compression via invertible image generation,

    F. Gao, X. Deng, J. Jing, X. Zou, and M. Xu, “Extremely low bit-rate image compression via invertible image generation,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 8, pp. 6993– 7004, 2023

    2023

  27. [27]

    Toward extreme image rescaling with generative prior and invertible prior,

    H. Wei, C. Ge, Z. Li, X. Qiao, and P. Deng, “Toward extreme image rescaling with generative prior and invertible prior,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 7, pp. 6181– 6193, 2024

    2024

  28. [28]

    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

  29. [29]

    Towards extreme image compression with latent feature guidance and diffusion prior,

    Z. Li, Y . Zhou, H. Wei, C. Ge, and J. Jiang, “Towards extreme image compression with latent feature guidance and diffusion prior,”IEEE Transactions on Circuits and Systems for Video Technology, 2024

    2024

  30. [30]

    Generative semantic communication: Diffusion models beyond bit recovery,

    E. Grassucci, S. Barbarossa, and D. Comminiello, “Generative semantic communication: Diffusion models beyond bit recovery,”arXiv preprint arXiv:2306.04321, 2023

    Pith/arXiv arXiv 2023

  31. [31]

    Semantic based generative compression of images for extremely low bitrates,

    T. Bordin and T. Maugey, “Semantic based generative compression of images for extremely low bitrates,” in2023 IEEE 25th International Workshop on Multimedia Signal Processing (MMSP). IEEE, 2023, pp. 1–6

    2023

  32. [32]

    Misc: Ultra-low bitrate image semantic compression driven by large multimodal model,

    C. Li, G. Lu, D. Feng, H. Wu, Z. Zhang, X. Liu, G. Zhai, W. Lin, and W. Zhang, “Misc: Ultra-low bitrate image semantic compression driven by large multimodal model,”IEEE Transactions on Image Processing, 2024

    2024

  33. [33]

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

    M. Careil, M. J. Muckley, J. Verbeek, and S. Lathuili`ere, “Towards image compression with perfect realism at ultra-low bitrates,” inThe Twelfth International Conference on Learning Representations, 2023

    2023

  34. [34]

    Rate-distortion optimized cross modal compression with multiple domains,

    J. Gao, C. Jia, Z. Huang, S. Wang, S. Ma, and W. Gao, “Rate-distortion optimized cross modal compression with multiple domains,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 8, pp. 6978–6992, 2024

    2024

  35. [35]

    Ultra lowrate image compression with semantic residual coding and compression-aware diffusion,

    A. Ke, X. Zhang, T. Chen, M. Lu, C. Zhou, J. Gu, and Z. Ma, “Ultra lowrate image compression with semantic residual coding and compression-aware diffusion,”arXiv preprint arXiv:2505.08281, 2025

    Pith/arXiv arXiv 2025

  36. [36]

    Exploring multimodal knowledge for image compression via large foundation models,

    J. Gao, Z. Huang, Q. Mao, S. Ma, and C. Jia, “Exploring multimodal knowledge for image compression via large foundation models,”IEEE Transactions on Image Processing, 2025

    2025

  37. [37]

    Turning a clip model into a scene text detector,

    W. Yu, Y . Liu, W. Hua, D. Jiang, B. Ren, and X. Bai, “Turning a clip model into a scene text detector,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 6978– 6988

    2023

  38. [38]

    A survey on multimodal large language models,

    S. Yin, C. Fu, S. Zhao, K. Li, X. Sun, T. Xu, and E. Chen, “A survey on multimodal large language models,”National Science Review, vol. 11, no. 12, p. nwae403, 2024

    2024

  39. [39]

    Adding conditional control to text-to-image diffusion models,

    L. Zhang, A. Rao, and M. Agrawala, “Adding conditional control to text-to-image diffusion models,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 3836–3847

    2023

  40. [40]

    Causal contextual prediction for learned image compression,

    Z. Guo, Z. Zhang, R. Feng, and Z. Chen, “Causal contextual prediction for learned image compression,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 4, pp. 2329–2341, 2021

    2021

  41. [41]

    Ensemble learning-based rate-distortion optimization for end-to-end image compression,

    Y . Wang, D. Liu, S. Ma, F. Wu, and W. Gao, “Ensemble learning-based rate-distortion optimization for end-to-end image compression,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 31, no. 3, pp. 1193–1207, 2020

    2020

  42. [42]

    A task-oriented real-time and robust feature compression and selection method in collaborative intelligence system,

    K. Wang, T. Zhao, and F. Li, “A task-oriented real-time and robust feature compression and selection method in collaborative intelligence system,” inICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  43. [43]

    Deep unsupervised learning using nonequilibrium thermodynamics,

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” in International conference on machine learning. pmlr, 2015, pp. 2256– 2265

    2015

  44. [44]

    Variational diffusion models,

    D. Kingma, T. Salimans, B. Poole, and J. Ho, “Variational diffusion models,”Advances in neural information processing systems, vol. 34, pp. 21 696–21 707, 2021

    2021

  45. [45]

    A resid- ual diffusion model for high perceptual quality codec augmentation,

    N. F. Ghouse, J. Petersen, A. Wiggers, T. Xu, and G. Sautiere, “A resid- ual diffusion model for high perceptual quality codec augmentation,” arXiv preprint arXiv:2301.05489, 2023

    Pith/arXiv arXiv 2023

  46. [46]

    What your features reveal: Data-efficient black-box feature inversion attack for split dnns,

    Z. Ren, L. He, J. Liang, X. Fu, H. Bi, and F. Li, “What your features reveal: Data-efficient black-box feature inversion attack for split dnns,” arXiv preprint arXiv:2511.15316, 2025

    arXiv 2025

  47. [47]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems, vol. 33, pp. 6840– 6851, 2020

    2020

  48. [48]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

    2022

  49. [49]

    Rfc 8878: Zstandard compression and the ’application/zstd’ media type,

    Y . Collet and M. Kucherawy, “Rfc 8878: Zstandard compression and the ’application/zstd’ media type,” USA, 2021. 14

    2021

  50. [50]

    Information entropy algorithms for image, video, and signal processing,

    G. Jeon, “Information entropy algorithms for image, video, and signal processing,” p. 926, 2021

    2021

  51. [51]

    An image is worth one word: Person- alizing text-to-image generation using textual inversion,

    R. Gal, Y . Alaluf, Y . Atzmon, O. Patashnik, A. H. Bermano, G. Chechik, and D. Cohen-Or, “An image is worth one word: Person- alizing text-to-image generation using textual inversion,”arXiv preprint arXiv:2208.01618, 2022

    Pith/arXiv arXiv 2022

  52. [52]

    Parameter-efficient transfer learning for nlp,

    N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. De Laroussilhe, A. Gesmundo, M. Attariyan, and S. Gelly, “Parameter-efficient transfer learning for nlp,” inInternational conference on machine learning. PMLR, 2019, pp. 2790–2799

    2019

  53. [53]

    Lora: Low-rank adaptation of large language models

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chenet al., “Lora: Low-rank adaptation of large language models.” ICLR, vol. 1, no. 2, p. 3, 2022

    2022

  54. [54]

    Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image cap- tioning,

    P. Sharma, N. Ding, S. Goodman, and R. Soricut, “Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image cap- tioning,” inProceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2018, pp. 2556– 2565

    2018

  55. [55]

    Testimages: a large-scale archive for testing visual devices and basic image processing algorithms

    N. Asuni, A. Giachettiet al., “Testimages: a large-scale archive for testing visual devices and basic image processing algorithms.” inSTAG, 2014, pp. 63–70

    2014

  56. [56]

    Clic 2020: Challenge on learned image compression,

    G. Toderici, L. Theis, N. Johnston, E. Agustsson, F. Mentzer, J. Ball ´e, W. Shi, and R. Timofte, “Clic 2020: Challenge on learned image compression,”Retrieved March, vol. 29, p. 2021, 2020

    2020

  57. [57]

    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,” inProceedings of the IEEE con- ference on computer vision and pattern recognition workshops, 2017, pp. 126–135

    2017

  58. [58]

    Image quality assessment: Unifying structure and texture similarity,

    K. Ding, K. Ma, S. Wang, and E. P. Simoncelli, “Image quality assessment: Unifying structure and texture similarity,”IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 5, pp. 2567– 2581, 2020

    2020

  59. [59]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 586–595

    2018

  60. [60]

    Semantic similarity score for measuring visual similarity at semantic level,

    S. Fan, Z. Bao, C. Dong, H. Liang, X. Xu, and P. Zhang, “Semantic similarity score for measuring visual similarity at semantic level,”IEEE Internet of Things Journal, 2024

    2024

  61. [61]

    Maniqa: Multi-dimension attention network for no-reference image quality assessment,

    S. Yang, T. Wu, S. Shi, S. Lao, Y . Gong, M. Cao, J. Wang, and Y . Yang, “Maniqa: Multi-dimension attention network for no-reference image quality assessment,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 1191–1200

    2022

  62. [62]

    Exploring clip for assessing the look and feel of images,

    J. Wang, K. C. Chan, and C. C. Loy, “Exploring clip for assessing the look and feel of images,” inProceedings of the AAAI conference on artificial intelligence, vol. 37, no. 2, 2023, pp. 2555–2563

    2023

  63. [63]

    Toward generalized image quality assessment: Relaxing the perfect reference quality assumption,

    D. Chen, T. Wu, K. Ma, and L. Zhang, “Toward generalized image quality assessment: Relaxing the perfect reference quality assumption,” inProceedings of the Computer Vision and Pattern Recognition Confer- ence, 2025, pp. 12 742–12 752

    2025

  64. [64]

    Multiscale structural similarity for image quality assessment,

    Z. Wang, E. P. Simoncelli, and A. C. Bovik, “Multiscale structural similarity for image quality assessment,” inThe thrity-seventh asilomar conference on signals, systems & computers, 2003, vol. 2. Ieee, 2003, pp. 1398–1402. Kaile Wangreceived the B.E. degree in telecommu- nications engineering from Xidian University, China, in 2024. He is currently pursui...

    2003

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.