Pith. sign in

REVIEW 5 major objections 6 minor 64 references

Conditional Latent Coding with Learnable Synthesized Reference for Deep Image Compression

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Conditional Latent Coding claims up to 1.2 dB gain in deep image compression with about 0.5% bit overhead.

desk verdict Plausible engineering result in a crowded subfield, but the theory section does not deliver the advertised dictionary-size bound and the decoder description is internally inconsistent. read the letter →

arxiv 2502.09971 v1 pith:BTBOJKXL submitted 2025-02-14 cs.CV cs.AI

classification cs.CVcs.AI
keywords deepimagecompressionconditionallatentcodingfeaturedictionaryreference-basedsynthesisrate-distortionoptimizationentropylearned
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 proposes Conditional Latent Coding (CLC), a deep image compression method that builds a universal image feature dictionary from a large set of external images and, for each input image, learns to synthesize a conditioning latent from retrieved dictionary features. The conditioning latent is used at both encoder and decoder, so the compressed representation can exploit correlations between the input and an external image collection rather than coding the image latent from scratch. The authors claim that this raises rate-distortion performance by up to 1.2 dB at low bitrates with an overhead of roughly 0.5% bits per pixel, and that the method tolerates imperfect dictionary retrieval because its error bound grows logarithmically with dictionary size. If the claims hold, existing learned codecs could be upgraded with a shared feature dictionary and a small per-image index cost.

What carries the argument

The load-bearing mechanism is the pair of modules that turn retrieved dictionary features into a condition for entropy coding. Conditional Latent Matching (CLM) computes a soft correspondence between the target latent $y$ and the reference latents $Y^M_r$ with a temperature-scaled softmax similarity, then aligns the matched features with deformable convolutions; Conditional Latent Synthesis (CLS) fuses the aligned reference $y_a$ into a Gaussian conditional model whose mean is $\mu(y,y_a)=\alpha \odot y+(1-\alpha)\odot y_a$, with $\alpha=\sigma(F_w([y,y_a]))$ computed by a small fusion network. The dictionary is built by modified spatial pyramid pooling, PCA, and MiniBatch K-means clustering into 3000 representatives, and retrieval is accelerated with a ball tree and a KV-cache with learned compression and eviction. A hyperprior plus slice-wise autoregressive entropy model codes the fused latent $y_f$. Together these components convert an external image collection into a compact side-information signal transmitted as dictionary indices and fused at both encoder and decoder.

What would settle it

Inspect the bitstream produced by the published code to see whether the original latent $y$ is entropy-coded separately; if it is, recompute the real overhead. Then decode test images using only the transmitted $y_f$, the dictionary indices, and the decoder's synthesis modules, and compare the recovered latent to the encoder's $y$; a large mismatch would show that the central recovery step is not actually performed.

Watch

Extended reading notes

Core claim

The central claim is that conditional latent coding with a learned, dynamically synthesized reference outperforms both traditional codecs and current learned codecs at low bitrates. CLC constructs a dictionary by extracting multi-scale features from reference images with a modified spatial pyramid pooling ResNet-50, reducing them with PCA, clustering them with MiniBatch K-means, and keeping one representative per cluster. For a target image, a ball-tree search and a KV-cache retrieve the top three matching references, and the CLM and CLS modules translate those references into a conditioning latent whose adaptive fusion with the target latent defines the coded representation $y_f$. The paper reports that this representation yields up to 1.2 dB gains and about 14.5% BD-rate savings over VTM on Kodak, and it derives a perturbation bound intended to show that retrieval error stays controlled as the dictionary grows.

Load-bearing premise

The load-bearing assumption is that the decoder can recover the image's own latent $y$ from the transmitted fused latent $y_f$ plus dictionary indices through the same matching and synthesis modules, since the paper specifies neither an invertible fusion step nor a separately coded copy of $y$; if that recovery is impossible, or if $y$ is actually sent, the claimed 0.5% overhead is not the true cost.

Editorial extensions

If this is right

  • At equal reconstruction quality on Kodak, CLC reports 0.1 bpp at MS-SSIM 0.95, where TCM, VTM, BPG, and JPEG need 0.15, 0.18, 0.22, and 0.38 bpp respectively.
  • The dictionary side information costs approximately 0.5% bits per pixel, because the main transmitted signal is the fused latent plus reference indices rather than a full per-image reference.
  • Three retrieved references give the best BD-rate savings on Kodak and CLIC; adding a fourth or fifth reference adds redundancy and slightly lowers the gain.
  • The perturbation analysis implies that moderate mismatches in dictionary retrieval cause only a small performance reduction, with degradation accelerating only when the perturbation rate grows large.
  • A dictionary cluster size around 3000 balances compression gains with encoding time; larger dictionaries multiply encoding time while adding only marginal BD-rate savings.

Reading between the lines

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

  • Editorial inference: if the decoder can indeed recover $y$ from $y_f$ and indices, the same dictionary could serve multiple images or video frames, turning CLC into a natural building block for conditional coding across time and views.
  • Editorial inference: the claimed 0.5% overhead is only the full cost if the original latent $y$ is not itself transmitted; checking the bitstream layout for a separately coded $y$ is the fastest way to test that assumption.
  • Editorial inference: the logarithmic dependence on dictionary size could be probed directly by measuring performance reduction at cluster sizes 1000, 2000, 4000, and 8000 and comparing the degradation curve with the stated bound.
  • Editorial inference: the conditioning latent idea should transfer to other learned codecs by inserting CLM and CLS ahead of their entropy models, but the gain may depend on how well the dictionary matches the target domain.
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

5 major / 6 minor

Summary. The paper proposes Conditional Latent Coding (CLC), a deep image compression method that constructs a universal feature dictionary via spatial pyramid pooling, PCA, and clustering; for each input image it retrieves top-M reference images, aligns and fuses their latents with the input latent through Conditional Latent Matching (CLM) and Conditional Latent Synthesis (CLS) modules, and entropy-codes the fused latent with a hyperprior and slice-based autoregressive model. The authors claim up to 1.2 dB improvement over existing codecs with roughly 0.5% bits-per-pixel overhead, and present a theoretical perturbation analysis asserting an error bound that scales logarithmically with dictionary size. The paper includes ablation studies, robustness experiments, and a public code release.

Significance. If the central claims hold, the method would offer a practical way to upgrade learned codecs with a shared feature dictionary and a small per-image index overhead, with meaningful rate-distortion gains at low bitrates. The paper has several genuine strengths: the full code is publicly released, the empirical comparisons are made against strong baselines (BPG, VTM, TCM, ELIC), and the ablation studies isolate component contributions. However, the significance is currently undercut by unresolved questions about what exactly is transmitted and reconstructed, and by a theoretical result that does not deliver the advertised dictionary-size dependence. These issues are central rather than cosmetic, so the paper requires substantial revision before its contribution can be assessed.

major comments (5)
  1. [§3.3(4) and Eq. (6)] The decoder-side reconstruction of y from ŷf is underspecified. Eq. (6) defines yf = α⊙y + (1−α)⊙ya, and ya depends on y through Eq. (4) and the alignment function Fa. Therefore y appears on both sides of the fusion equation, and no inverse operation is given in the main text. Algorithm 2 does not resolve this: it reconstructs ŷf from the hyperprior z alone (line 3), then entropy-decodes slices of ŷ (lines 14–18), which is a different codec in which y (or its slices) is transmitted. If y is coded separately, the claimed ~0.5% bpp overhead is not the full side-information cost; if y is not coded, the decoder is not defined. This is the load-bearing issue for the paper's central compression claim.
  2. [Abstract and §2 (Unique Contributions)] The abstract and Section 2 state that the theoretical error bound scales logarithmically with the dictionary size. However, neither Theorem 1 in the main text nor Theorem 2 in Appendix A.3 contains any dictionary-size parameter; the logarithmic term is log(d/δ), where d is the ambient signal dimension. The advertised dictionary-size dependence is therefore not established by the provided theory.
  3. [Appendix A.3, Theorem 2 and Remark 1] The proof's Step 5 introduces δeff = λmin(Σs)(1−ρ) − σξ² − σξ̃² without derivation from the model in Assumptions 1–2. Moreover, the resulting bound is internally inconsistent with the model's semantics: as ρ→1 (highly correlated reference, irrelevant proportion p = 1−ρ² → 0), the denominator (1−ρ) tends to zero and the bound grows worse, directly contradicting the intuition stated in Remark 2 and Remark 3. This sign issue makes the theoretical interpretation untenable as written.
  4. [Appendix A.3, Steps 3–4] The eigenvalue-gap argument is asserted rather than derived. The claimed gap δgap = λmin(Σs) − σξ² is stated without justification, and the Davis–Kahan application in Eq. (25) with the sampling-error bound of Eq. (22) is not carried out. In addition, the main-text Theorem 1 and the appendix Theorem 2 are different statements (the former includes a min and the effective rank r(Σξ), the latter does not), and the relationship between them is unexplained.
  5. [§4.2, Figure 4, and Tables 1–3] The headline quantitative claims of "up to 1.2 dB" improvement and "approximately 0.5% bits per pixel" overhead are not supported by any numerical table, error bar, or direct measurement of the side-information bitrate. The ablation tables report BD-rate savings relative to VTM, but the actual bitstream composition—how many bits are spent on dictionary indices versus latent code slices—is never reported. Without this breakdown, the central compression-efficiency claim cannot be verified from the manuscript.
minor comments (6)
  1. [§3.3(3)] The citation for the slice-based autoregressive context model is empty: the text reads "()" instead of a reference.
  2. [§3.2–§3.3] The symbol K is used for three different quantities: the number of dictionary clusters, the number of slices in the autoregressive model, and the key dimension in the KV-cache. Please disambiguate these to avoid confusion.
  3. [Algorithms 1 and 2] Algorithm 1 computes residuals and distribution parameters but never shows the actual entropy-coding of the quantized slices, while Algorithm 2 line 3 writes "ŷf ← h_a^{-1}(z)", implying that the hyperprior z alone determines ŷf. This is inconsistent with z being a hyperprior side-information representation and needs clarification.
  4. [Theorem 1 vs. Theorem 2] The main-text Theorem 1 and the appendix Theorem 2 are stated in different forms and with different parameters; the paper should state one theorem or explicitly explain the equivalence.
  5. [Appendix B, Figure 6] The axis labels and surrounding text of Figure 6 appear garbled in the manuscript, making the robustness results unreadable.
  6. [§3.2 vs. §4.1] Section 3.2 says the dictionary is built from "randomly download[ed] 3000 images from the web," while Section 4.1 says the Flickr2K dataset (2650 images) is used for dictionary construction; these statements should be reconciled.

Circularity Check

1 steps flagged · score 4.0 of 10

Decoder reconstruction is self-referential: Eq. (6) defines yf from y, yet §3.3(4) claims y is recovered from yf by the same procedure, and Algorithm 2 contradicts this by entropy-decoding y itself.

  1. self definitional [Section 3.3(4) 'Decoding and Optimization', Eqs. (4)–(6), Algorithm 2]
    "During decoding, we first reconstruct ˆz and ˆyf from the bitstream. Then, using the dictionary indices passed from the encoder, we apply the same reference processing and alignment procedure to reconstruct y from ˆyf. ... µ(y, ya) = α ⊙ y + (1 − α) ⊙ ya, (6) ... ya = Fa(y, ym; θa) ... Sij = exp(⟨ϕ(yi), ϕ(yr,j)⟩/τ )P k exp(⟨ϕ(yi), ϕ(yr,k)⟩/τ ) , (4)"

    The transmitted quantity yf is defined by Eqs. (5)–(6) as a function of y itself: ya is computed from y via CLM (Eq. 4) and CLS (Fa), and the fusion mean is µ(y, ya) = α⊙y + (1−α)⊙ya with α = σ(Fw([y, ya])). Therefore 'reconstruct y from yf using the same reference processing and alignment procedure' would require y as an input to the very matching/alignment/fusion steps that are supposed to output y; no inverse of the fusion or additional conditioning signal is specified. Algorithm 2 side-steps this by decoding slices of y directly from the bitstream, which would make y part of the transmitted rate and contradict the claimed ~0.5% bpp overhead.

full rationale

The central empirical claim—up to 1.2 dB improvement with about 0.5% bpp overhead—is benchmarked against external codecs (VTM, BPG, TCM, ELIC, etc.) rather than fitted to the test sets, so the RD comparison itself is not circular. The theoretical robustness result is an independent Davis-Kahan/Matrix Bernstein argument under a spiked covariance model; it does not assume the CLC architecture or feed back into the training objective, and therefore is not a self-citation or a restatement of the experimental results. However, the method's decoding chain contains a genuine self-reference: yf is defined as a function of y in Eq. (6), yet §3.3(4) says y is reconstructed from yf by the same procedure, and Algorithm 2 resolves the contradiction only by entropy-decoding y slices, which would change the rate accounting. This makes a load-bearing part of the derivation self-referential or underspecified, though the published RD numbers are externally anchored. Hence a moderate circularity score of 4 is appropriate: partial circularity in the method description, but the main benchmark claim retains independent content.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The central claim rests on several hand-selected hyperparameters, a theoretical model that does not match the actual retrieval mechanism, and an implicit decoder-side invertibility assumption. The most load-bearing items are the CLS inversion assumption and the unproven eigenvalue-gap reduction.

free parameters (5)
  • M, number of reference images = 3
    Ablation Table 1 selects M=3 as best BD-rate on Kodak and CLIC; all headline BD-rate numbers use this value.
  • Dictionary cluster count K = 3000
    Ablation Table 2 picks 3000 as the best trade-off between BD-rate and encoding time; the dictionary and retrieval depend on this choice.
  • Number of slices in autoregressive context model = 8
    Implementation detail in Section 4.1; changes probability estimation and rate, with no sensitivity analysis reported.
  • KV-cache size N = 300, adjusted by GPU memory
    Section 4.1; affects retrieval diversity and encoding time, and it is not ablated.
  • Initial fusion weight alpha_0 = 0.5
    Section 4.1; initial value for the learned fusion weight, likely minor after training.
assumptions (5)
  • domain assumption Natural images follow the spiked covariance model x = U*s + xi with sub-Gaussian noise.
    Assumptions 1 and 5 in Appendix A; no empirical validation, though such a model is common in high-dimensional statistics.
  • ad hoc to paper The reference image model x_tilde = U*(rho*s + sqrt(1-rho^2)*s_perp) + xi_tilde captures how dictionary references relate to the input.
    Assumption 2; specific to this paper's theory, and rho is not estimated from the actual dictionary retrieval.
  • ad hoc to paper Retrieved reference images reduce the effective eigenvalue gap to delta_eff = lambda_min(Sigma_s)*(1 - rho) - sigma_xi^2 - sigma_xi_tilde^2.
    Appendix A Step 5; stated without derivation and is the mechanism that produces the theorem's rho dependence.
  • domain assumption The CLS fusion y_f = alpha * y + (1 - alpha) * y_a is invertible enough on the decoder to recover y from y_f and references.
    Section 3.3 decoding paragraph; no inverse network or transmitted alpha is specified.
  • standard math Matrix Bernstein and Davis-Kahan inequalities hold under the assumed noise.
    Cited in Appendix A Steps 2 and 4; accepted background, but applied to a covariance matrix where the reference image is not present.
invented entities (1)
  • Conditioning latent y_f
    purpose: A per-image fused reference latent that guides entropy coding and reconstruction.
    It is an internal learned representation with no observable handle outside the network, so its usefulness is only shown by the reported codec results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Conditional Latent Coding with Learnable Synthesized Reference for Deep Image Compression." pith.science (2026). https://pith.science/paper/BTBOJKXL

@misc{pith2026250209971,
  author       = {Pith},
  title        = {Pith review of: Conditional Latent Coding with Learnable Synthesized Reference for Deep Image Compression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BTBOJKXL}},
  note         = {Machine review of arXiv:2502.09971}
}
read the original abstract

In this paper, we study how to synthesize a dynamic reference from an external dictionary to perform conditional coding of the input image in the latent domain and how to learn the conditional latent synthesis and coding modules in an end-to-end manner. Our approach begins by constructing a universal image feature dictionary using a multi-stage approach involving modified spatial pyramid pooling, dimension reduction, and multi-scale feature clustering. For each input image, we learn to synthesize a conditioning latent by selecting and synthesizing relevant features from the dictionary, which significantly enhances the model's capability in capturing and exploring image source correlation. This conditional latent synthesis involves a correlation-based feature matching and alignment strategy, comprising a Conditional Latent Matching (CLM) module and a Conditional Latent Synthesis (CLS) module. The synthesized latent is then used to guide the encoding process, allowing for more efficient compression by exploiting the correlation between the input image and the reference dictionary. According to our theoretical analysis, the proposed conditional latent coding (CLC) method is robust to perturbations in the external dictionary samples and the selected conditioning latent, with an error bound that scales logarithmically with the dictionary size, ensuring stability even with large and diverse dictionaries. Experimental results on benchmark datasets show that our new method improves the coding performance by a large margin (up to 1.2 dB) with a very small overhead of approximately 0.5\% bits per pixel. Our code is publicly available at https://github.com/ydchen0806/CLC.

Figures

Figures reproduced from arXiv: 2502.09971 by the authors.

Figure 1
Figure 1. Overview of the proposed Conditional Latent Cod [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Universal Feature Dictionary Construction. (a) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The detail of our proposed CLM and CLS module. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: The rate-distortion performance comparison of different methods. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Image reconstruction results at around 0.1 bpp. From left to right: Raw inputs, reference images, reconstructed images. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Performance Reduction (PR) of CLC under vary [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Visualization of reconstructed images using our method. The red and blue boxes highlight magnified areas for detailed [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Visualization of reconstructed images using our method. The red and blue boxes highlight magnified areas for detailed [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

64 extracted references · 55 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Ayzik, S.; and Avidan, S. 2020. Deep image compression using decoder side information. In ECCV, 699--714

  4. [4]

    Ball \'e , J.; Laparra, V.; and Simoncelli, E. P. 2017. End-to-end optimized image compression. In ICLR

  5. [5]

    J.; and Johnston, N

    Ball \'e , J.; Minnen, D.; Singh, S.; Hwang, S. J.; and Johnston, N. 2018. Variational image compression with a scale hyperprior. In ICLR

  6. [6]

    Bellard, F. 2014. BPG Image Format. https://bellard.org/bpg/

  7. [7]

    J.; and Wang, Y.-K

    Bross, B.; Chen, J.; Ohm, J.-R.; Sullivan, G. J.; and Wang, Y.-K. 2021. Overview of the Versatile Video Coding (VVC) Standard and its Applications. IEEE Transactions on Circuits and Systems for Video Technology

  8. [8]

    Chen, Y.; Huang, W.; Liu, X.; Deng, S.; Chen, Q.; and Xiong, Z. 2024 a . Learning multiscale consistency for self-supervised electron microscopy instance segmentation. In ICASSP

Show all 64 references
  1. [9]

    Chen, Y.; Huang, W.; Zhou, S.; Chen, Q.; and Xiong, Z. 2023 a . Self-supervised neuron segmentation with multi-agent reinforcement learning. In IJCAI

  2. [10]

    Chen, Y.; Liu, C.; Huang, W.; Cheng, S.; Arcucci, R.; and Xiong, Z. 2023 b . Generative text-guided 3d vision-language pretraining for unified medical image segmentation. arXiv preprint arXiv:2306.04811

  3. [11]

    Chen, Y.; Liu, C.; Liu, X.; Arcucci, R.; and Xiong, Z. 2024 b . BIMCV-r: A landmark dataset for 3d ct text-image retrieval. In MICCAI

  4. [12]

    Chen, Y.; Mukherjee, D.; Han, J.; Grange, A.; Xu, Y.; Liu, Z.; Parker, S.; Chen, C.; Agarwal, H.; Deshpande, S.; et al. 2018. An Overview of Core Coding Tools in the AV1 Video Codec. In PCS. IEEE

  5. [13]

    Chen, Y.; Shi, H.; Liu, X.; Shi, T.; Zhang, R.; Liu, D.; Xiong, Z.; and Wu, F. 2024 c . TokenUnify: Scalable Autoregressive Visual Pre-training with Mixture Token Prediction. arXiv preprint arXiv:2405.16847

  6. [14]

    Chen, Z.; Wang, R.; He, D.; Zhang, L.; and Ma, S. 2023 c . Transformer-based Context Modeling for Image Compression. In CVPR

  7. [15]

    Cheng, Z.; Sun, H.; Takeuchi, M.; and Katto, J. 2020. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In CVPR, 7939--7948

  8. [16]

    Cheng, Z.; Sun, H.; Takeuchi, M.; and Katto, J. 2021. Learned Image Compression with Discretized Gaussian Mixture Likelihoods and Attention Modules. In CVPR, 7939--7948

  9. [17]

    Deng, S.; Chen, Y.; Huang, W.; Zhang, R.; and Xiong, Z. 2024. Unsupervised Domain Adaptation for EM Image Denoising with Invertible Networks. IEEE Transactions on Medical Imaging

  10. [18]

    Hu, D.; Dong, Z.; Liang, K.; Yu, H.; Wang, S.; and Liu, X. 2024 a . High-order Topology for Deep Single-cell Multi-view Fuzzy Clustering. IEEE Transactions on Fuzzy Systems

  11. [19]

    Hu, D.; Liu, S.; Wang, J.; Zhang, J.; Wang, S.; Hu, X.; Zhu, X.; Tang, C.; and Liu, X. 2024 b . Reliable Attribute-missing Multi-view Clustering with Instance-level and feature-level Cooperative Imputation. In Proceedings of the 32nd ACM International Conference on Multimedia,...

  12. [20]

    Huang, Y.; Chen, B.; Qin, S.; Li, J.; Wang, Y.; Dai, T.; and Xia, S.-T. 2023. Learned distributed image compression with multi-scale patch matching in feature domain. In AAAI, volume 37, 4322--4329

  13. [21]

    Jia, Z.; Li, J.; Li, B.; Li, H.; and Lu, Y. 2024. Generative Latent Coding for Ultra-Low Bitrate Image Compression. In CVPR, 26088--26098

  14. [22]

    Kodak, E. 1993. Kodak Lossless True Color Image Suite (PhotoCD PCD0992). Version 5

  15. [23]

    Lee, J.; Cho, S.; and Beack, S.-K. 2019. Context-adaptive Entropy Model for End-to-end Optimized Image Compression. In ICLR

  16. [24]

    Li, J.; Li, B.; and Lu, Y. 2021. Deep contextual video compression. In NeurIPS, volume 34, 18114--18125

  17. [25]

    Li, J.; Li, B.; and Lu, Y. 2023. Neural video compression with diverse contexts. In CVPR, 22616--22626

  18. [26]

    Li, L.; Xing, J.; Yu, X.; and Zhang, X.-P. 2024 a . Deviation Wing Loss for High-Performance 2D Pose Estimation. In 2024 IEEE International Conference on Multimedia and Expo (ICME), 1--6. IEEE

  19. [27]

    Li, L.; Yang, W.; Yu, X.; Xing, J.; and Zhang, X.-P. 2024 b . Translating Motion to Notation: Hand Labanotation for Intuitive and Comprehensive Hand Movement Documentation. In Proceedings of the 32nd ACM International Conference on Multimedia, 4092--4100

  20. [28]

    Li, M.; Shen, L.; Ye, P.; Feng, G.; and Wang, Z. 2023. RFD-ECNet: Extreme Underwater Image Compression with Reference to Feature Dictionary. In ICCV, 12980--12989

  21. [29]

    C.; Ma, L.; Fu, J.; Guo, Y.; Shah, A.; Bai, W.; and Arcucci, R

    Liu, C.; Ouyang, C.; Chen, Y.; Quilodr \'a n-Casas, C. C.; Ma, L.; Fu, J.; Guo, Y.; Shah, A.; Bai, W.; and Arcucci, R. 2023. T3d: Towards 3d medical image understanding through vision-language pre-training. arXiv preprint arXiv:2312.01529

  22. [30]

    Liu, J.; Lu, G.; Hu, Z.; and Xu, D. 2020. A unified end-to-end framework for efficient deep image compression. arXiv preprint arXiv:2002.03370

  23. [31]

    Liu, J.; Sun, H.; and Katto, J. 2023. Learned image compression with mixed transformer-cnn architectures. In CVPR, 14388--14397

  24. [32]

    Liu, X.; Cai, M.; Chen, Y.; Zhang, Y.; Shi, T.; Zhang, R.; Chen, X.; and Xiong, Z. 2024. Cross-dimension affinity distillation for 3d em neuron segmentation. In CVPR

  25. [33]

    Ma, X.; Lian, R.; Wu, Z.; Guo, H.; Ma, M.; Wu, S.; Du, Z.; Song, S.; and Zhang, W. 2024. LOGCAN++: Adaptive Local-global class-aware network for semantic segmentation of remote sensing imagery. arXiv:2406.16502

  26. [34]

    Ma, X.; Ma, M.; Hu, C.; Song, Z.; Zhao, Z.; Feng, T.; and Zhang, W. 2023. Log-Can: Local-Global Class-Aware Network For Semantic Segmentation of Remote Sensing Images. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 1--5

  27. [35]

    Mentzer, F.; Agustsson, E.; Tschannen, M.; Timofte, R.; and Van Gool, L. 2018. Conditional probability models for deep image compression. In CVPR, 4394--4402

  28. [36]

    Minnen, D.; Ballé, J.; and Toderici, G. D. 2018. Joint autoregressive and hierarchical priors for learned image compression. In NeurIPS, 10771--10780

  29. [37]

    Qian, H.; Chen, Y.; Lou, S.; Khan, F.; Jin, X.; and Fan, D.-P. 2024. Maskfactory: Towards high-quality synthetic data generation for dichotomous image segmentation. In NeurIPS

  30. [38]

    Qian, Y.; Lin, M.; Sun, X.; Tan, Z.; and Jin, R. 2022. Entroformer: A transformer-based entropy model for learned image compression. arXiv preprint arXiv:2202.05492

  31. [39]

    Shen, F.; and Tang, J. 2024. IMAGPose: A Unified Conditional Framework for Pose-Guided Person Generation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

  32. [40]

    Shen, F.; Ye, H.; Zhang, J.; Wang, C.; Han, X.; and Wei, Y. 2024. Advancing Pose-Guided Image Synthesis with Progressive Conditional Diffusion Models. In The Twelfth International Conference on Learning Representations

  33. [41]

    Sheng, X.; Li, J.; Li, B.; Li, L.; Liu, D.; and Lu, Y. 2022. Temporal context mining for learned video compression. IEEE Transactions on Multimedia, 25: 7311--7322

  34. [42]

    J.; Ohm, J.-R.; Han, W.-J.; and Wiegand, T

    Sullivan, G. J.; Ohm, J.-R.; Han, W.-J.; and Wiegand, T. 2012. Overview of the High Efficiency Video Coding (HEVC) Standard. IEEE Transactions on Circuits and Systems for Video Technology, 22(12): 1649--1668

  35. [43]

    Sun, H. 2024. Ultra-High Resolution Segmentation via Boundary-Enhanced Patch-Merging Transformer. arXiv:2412.10181

  36. [44]

    Sun, H.; Xu, L.; Jin, S.; Luo, P.; Qian, C.; and Liu, W. 2024. PROGRAM: PROtotype GRAph Model based Pseudo-Label Learning for Test-Time Adaptation. In The Twelfth International Conference on Learning Representations

  37. [45]

    Sun, J.; Zheng, N.-N.; Tao, H.; and Shun, H.-Y. 2003. Image hallucination with primal sketch priors. In CVPR

  38. [46]

    Tao, H.; Li, J.; Hua, Z.; and Zhang, F. 2023. DUDB: Deep Unfolding Based Dual-Branch Feature Fusion Network for Pan-sharpening remote sensing images. IEEE Transactions on Geoscience and Remote Sensing

  39. [47]

    Timofte, R.; Agustsson, E.; Van Gool, L.; Yang, M.-H.; and Zhang, L. 2017. Ntire 2017 challenge on single image super-resolution: Methods and results. In CVPR workshops, 114--125

  40. [48]

    Toderici, G.; Shi, W.; Timofte, R.; Theis, L.; Ball \'e , J.; Agustsson, E.; Johnston, N.; and Mentzer, F. 2020. CLIC: Workshop and challenge on learned image compression. In CVPR workshop

  41. [49]

    Toderici, G.; Vincent, D.; Johnston, N.; Jin Hwang, S.; Minnen, D.; Shor, J.; and Covell, M. 2017. Full resolution image compression with recurrent neural networks. In CVPR, 5306--5314

  42. [50]

    Wallace, G. K. 1992. The JPEG still picture compression standard. IEEE Transactions on Consumer Electronics, 38(1): xviii--xxxiv

  43. [51]

    Wu, Y.; Chen, Z.; Wen, Y.; Huang, H.; and Yuan, H. 2022. Content-aware reference frame selection for efficient video compression. IEEE Transactions on Image Processing, 31: 5186--5198

  44. [52]

    Xiong, Z.; Sun, X.; and Wu, F. 2010. Robust web image/video super-resolution. IEEE transactions on image processing, 19(8): 2017--2028

  45. [53]

    Yang, R. 2023. Tinc: Tree-structured implicit neural compression. In CVPR, 18517--18526

  46. [54]

    Yang, R.; Chen, Y.; Zhang, Z.; Liu, X.; Li, Z.; He, K.; Xiong, Z.; Suo, J.; and Dai, Q. 2024. UniCompress: Enhancing Multi-Data Medical Image Compression with Knowledge Distillation. arXiv preprint arXiv:2405.16850

  47. [55]

    Yin, J.; Yan, S.; Chen, T.; Chen, Y.; and Yao, Y. 2024. Class Probability Space Regularization for semi-supervised semantic segmentation. Computer Vision and Image Understanding, 104146

  48. [56]

    Yuan, Z.; Cao, J.; Li, Z.; Jiang, H.; and Wang, Z. 2024 a . SD-MVS : Segmentation-Driven Deformation Multi-View Stereo with Spherical Refinement and EM Optimization . In Proceedings of the AAAI Conference on Artificial Intelligence , volume 38, 6871--6880

  49. [57]

    Yuan, Z.; Liu, C.; Shen, F.; Li, Z.; Luo, J.; Mao, T.; and Wang, Z. 2024 b . MSP-MVS : Multi-Granularity Segmentation Prior Guided Multi-View Stereo . arXiv:2407.19323

  50. [58]

    Yuan, Z.; Luo, J.; Shen, F.; Li, Z.; Liu, C.; Mao, T.; and Wang, Z. 2024 c . DVP-MVS : Synergize Depth-Edge and Visibility Prior for Multi-View Stereo . arXiv:2412.11578

  51. [59]

    Yue, H.; Sun, X.; Yang, J.; and Wu, F. 2013. Cloud-based image coding for mobile devices—Toward thousands to one compression. IEEE transactions on multimedia, 15(4): 845--857

  52. [60]

    Zhang, D.; Chen, D.; Zhi, P.; Chen, Y.; Yuan, Z.; Li, C.; Sunjing; Zhou, R.; and Zhou, Q. 2024. MapExpert: Online HD Map Construction with Simple and Efficient Sparse Map Element Expert. arXiv:2412.12704

  53. [61]

    Zhang, D.; Zhi, P.; Yong, B.; Wang, J.-Q.; Hou, Y.; Guo, L.; Zhou, Q.; and Zhou, R. 2023. EHSS: An Efficient Hybrid-supervised Symmetric Stereo Matching Network. 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC), 1044--1051

  54. [62]

    Zhao, J.; Li, B.; Li, J.; Xiong, R.; and Lu, Y. 2021. A universal encoder rate distortion optimization framework for learned compression. In CVPR, 1880--1884

  55. [63]

    Zhao, J.; Li, B.; Li, J.; Xiong, R.; and Lu, Y. 2023. A universal optimization framework for learning-based image codec. ACM Transactions on Multimedia Computing, Communications and Applications, 20(1): 1--19

  56. [64]

    Zou, F.; Feng, Y.; Wei, Y.; and Ren, J. 2022. ELIC: Efficient Learned Image Compression with Unevenly Grouped Space-Channel Contextual Adaptive Coding. In CVPR, 5718--5727

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.