Pith. sign in

REVIEW 3 major objections 5 minor 33 references

Your Demands Deserve More Bits: Referring Semantic Image Compression at Ultra-low Bitrate

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A new framework, Referring Semantic Image Compression (RSIC), lets a user's referring word control where bits are spent at ultra-low bitrates, improving fidelity in the focus area while preserving global realism and semantic alignment…

desk verdict A genuinely new user-guided semantic compression system, but the paper never isolates the guidance gradient from the extra bits it spends, so its headline fidelity claim is not yet pinned down. read the letter →

arxiv 2506.00526 v1 pith:H5WAYJJR submitted 2025-05-31 eess.IV

Chenhao Wu , Qingbo Wu , Haoran Wei , Shuai Chen , Mingzhou He , King Ngi Ngan , Fanman Meng , Hongliang Li This is my paper
classification eess.IV
keywords learnedimagecompressionreferringsemanticultra-lowbitratediffusionmodelspatiallyvariable-ratecodinguser-guideddecodingDDIMinversionfidelity
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

This paper argues that the main reliability problem of ultra-low-bitrate semantic image compression — that the same text description can decode into a different object instance — can be solved by letting the user say which part of the image matters. It introduces Referring Semantic Image Compression (RSIC), which sends a global text description, a small spatial weight map derived from a referring word, and a compressed latent code that spends more bits in the weighted region. The decoder then conditions a pretrained diffusion model on the global description while steering it toward the encoded focus region through per-step gradient guidance. The paper's claim is that this referring mode improves fidelity in the user-specified area, measured by f-PSNR and f-LPIPS, while keeping global realism (FID, KID) and semantic alignment (CLIP Score) competitive with the state-of-the-art PerCo at bitrates around 0.002–0.03 bpp. If true, it would make extreme-ratio semantic compression practical for applications that need faithful reconstruction of user-chosen content.

What carries the argument

The load-bearing machinery is a pair of modules: Hierarchical Spatially Variable-rate Latent Coding (HSVLC), a four-scale latent codec with M-gated entropy models and spatial feature transforms that allocates bits according to the user weight map M; and Guided Generative Decoding (GGD), which uses DDIM inversion to turn the compressed latent $\hat{z}_0$ into inverted latents $\hat{z}_t$ and then performs denoising with an extra guidance gradient $\gamma \nabla \| C(\tilde{z}_t) - \hat{z}_t \|^2$, combined with classifier-free global conditioning and per-step self-recurrence noise re-injection. The guidance formula rests on the identity $\nabla \log p(\tilde{z}_t \mid \hat{z}_0) = \nabla \log p(\tilde{z}_t \mid \hat{z}_t)$, which the paper uses to justify steering each denoising step toward the inverted latent. Together the two modules convert a referring word into both a bit-allocation policy and a decoding steering signal.

What would settle it

A concrete way to test the central claim is to run RSIC in referring mode with the guidance gradient disabled by setting gamma to zero while keeping the same M-gated bit allocation, and compare f-PSNR in the focus region. If the fidelity gain over global mode does not shrink, the improvement is due to spatial bit allocation alone, not the guided decoding step; if it vanishes, the per-step guidance term is doing the work claimed. A second check is to apply the referring map to a region with no salient object, such as open sky: if the guidance still produces large f-PSNR gains, the method is responding to location rather than semantic content.

Watch

Extended reading notes

Core claim

The central claim of the paper is that user demands can be injected into a pretrained generative codec without joint fine-tuning. RSIC encodes a global description with a vision-language model, encodes a spatially non-uniform latent feature whose bit allocation is gated by a user weight map M, and decodes by running DDIM inversion on the compressed latent, then denoising under classifier-free guidance from the global description while adding a per-step fidelity term $\gamma \nabla \| C(\tilde{z}_t) - \hat{z}_t \|^2$ that pulls each latent toward the codec reconstruction of the original at the same diffusion time. The reported result is that in referring mode the focus area gains fidelity — higher f-PSNR, lower f-LPIPS, and better f-SSIM — compared with global mode at similar bitrates, while global realism and semantic alignment stay on par with PerCo. The ablation study supports the attribution: disabling the referring guidance raises f-LPIPS from 0.2505 to 0.6253, and disabling the global description cuts CLIP Score from about 30.3 to 26.8.

Load-bearing premise

The decisive assumption is that the compressed version of the original image, after being inverted back through the denoising process, can stand in for the original well enough that pushing each generation step toward it is the same as conditioning the generation on the original; if this equivalence fails, the reported gains in the focus region may be artifacts of a hand-tuned correction step rather than genuine conditional generation.

Editorial extensions

If this is right

  • At ultra-low bitrates (around 0.002–0.03 bpp), a user can name an object and the decoded image will keep that object much closer to the original, measured by f-PSNR and f-LPIPS, while the rest of the image is generated from the global description.
  • Because HSVLC spends bits according to the same weight map that guides decoding, the method converts a linguistic demand directly into a rate-allocation policy, so the extra bits go where the user asked and the referring map itself costs at most about 0.0007 bpp.
  • Since no diffusion-model weights are fine-tuned, RSIC is compatible with any off-the-shelf text-to-image diffusion backbone, so improvements in base generative models could transfer directly to the compression pipeline.
  • The comparison between global and referring modes on MS-COCO, Kodak, and CLIC indicates that the same framework covers both a uniform 'keep everything plausible' mode and a targeted 'keep this object faithful' mode, with the user weight map as the only switch.

Reading between the lines

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

  • Because the guidance strength $\gamma$ is set by a fixed hand-picked schedule rather than learned, the method's Pareto frontier between local fidelity and global realism likely depends on $\gamma$; a reader could test whether adapting $\gamma$ per region or per step widens the operating range beyond the reported settings.
  • The equality $\nabla \log p(\tilde{z}_t \mid \hat{z}_0) = \nabla \log p(\tilde{z}_t \mid \hat{z}_t)$ is the unproven hinge; if DDIM inversion is imperfect, the guided trajectory may be pulled toward a pseudo-latent that does not lie on the real image manifold, which would make the realism-versus-fidelity balance deteriorate at higher guidance strengths.
  • The framework naturally extends to multiple demands by combining several referring maps; one would expect that per-step self-recurrence would need more recurrence steps to satisfy multiple simultaneous region constraints, which is a testable extension of the current single-region design.
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

3 major / 5 minor

Summary. The paper proposes RSIC, a referring semantic image compression framework for ultra-low bitrates. The encoder produces a LLaVA-based global text description and a spatially non-uniform referring latent code (HSVLC) gated by a user-derived weight map M; the decoder runs a pretrained Stable Diffusion model with classifier-free guidance plus a per-step gradient guidance term that pushes noisy latents toward the decoded latent. The authors claim this flexibly balances local fidelity, global realism, semantic alignment, and bitrate, and they present comparisons on MS-COCO, Kodak, and CLIC showing improved foreground fidelity in referring mode while matching PerCo on semantic and realism metrics.

Significance. The referring-compression idea is practically motivated, and the plug-and-play design (no diffusion fine-tuning) is a clear strength. The paper reports a broad set of metrics on three datasets and provides a pseudocode-level description of the method. However, the theoretical justification in Section II.C is a heuristic chain of equalities, and the ablation study does not isolate the gradient guidance term that is claimed to produce the local-fidelity gains. As presented, the empirical evidence supports an effect of the referring bit allocation more strongly than it supports the guided-decoding mechanism.

major comments (3)
  1. [Section II.C, Eq. (3)] The derivation equates ∇ log p(z̃t|ẑ0) with ∇ log p(z̃t|ẑt) and then with ∇ log p(z̃t) + ∇ log p(ẑt|z̃t). The first equality is not a consequence of Bayes' theorem; it holds only if the DDIM-inverted variable ẑt is a sufficient statistic for ẑ0, which is not established and is only approximate for the finite-step inversion used in Algorithm 1. The second replacement, in which ∇ log p(ẑt|z̃t) is treated as γ∇∥C(z̃t)−ẑt∥², additionally assumes that the MSE distortion is an unnormalized log-likelihood. These are substantive assumptions rather than deductions; the paper should either prove them or explicitly label them as heuristics, because the gradient term is the component that distinguishes referring mode from global mode.
  2. [Section II.C, Algorithm 1] The guidance term γ∇∥C(z̃t)−ẑt∥² is evaluated on noisy latents z̃t, but HSVLC is trained to encode clean latents z0; at large t the input to C is far out of distribution, so the interpretation of the gradient as a likelihood score is questionable. The scale γ=10^3√(ᾱ_{t−1}/ᾱ_t) is selected without sensitivity analysis. Because the central claim is that guided generative decoding improves local fidelity, the paper needs a direct validation of this term, for example an ablation with γ=0 at the same referring bit allocation and ideally a sweep over γ.
  3. [Section III.D, Table I] The ablation 'w/o RGE' removes both the extra referring bits and the guidance gradient, so it cannot separate the effect of GGD from the effect of HSVLC's M-gated bit allocation. Since DDIM inversion starts from ẑ0, the extra local detail encoded by HSVLC is already present in ẑ_T and would propagate through ordinary DDIM sampling even with γ=0. The reported f-PSNR and f-LPIPS improvement of referring mode over global mode could therefore be entirely a rate-allocation effect. Add a row that keeps the referring-mode ẑ0 and the same bitrate allocation while setting γ=0; this is the minimal experiment that tests the paper's central mechanism.
minor comments (5)
  1. [Sections I and II] There are several typos: 'informtion' in Section I, 'recieving' in Section II.A, 'boudaries' in Section II.B, 'denosing' in Section II.C, and 'acounts' in Section III.D.
  2. [Algorithm 1] In step 8, the gradient is written without a subscript; it should be ∇_z̃t to make clear with respect to which variable the derivative is taken.
  3. [Section III.A] The training setting 'λ = 0.01 exp7M' is ambiguous; please clarify whether this means λ = 0.01·exp(7M) and how the scalar λ depends on the spatial map M.
  4. [Section II.B] The claim that M consumes a maximum of 0.0007 bpp should be stated more carefully: with 1/64 spatial resolution and 3 bits per entry, the cost is 3/4096 ≈ 0.00073 bpp regardless of image resolution, so the reported value is accurate but the derivation should be shown.
  5. [Section IV] The paper does not discuss failure cases, such as what happens when the user-specified referring word is absent from the image or when the generated weight map M is inaccurate; a short limitations paragraph would improve the presentation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the central claim is empirical and the guidance term is a modeling choice, not a hidden reuse of the evaluation metric.

full rationale

The paper's central claim is an empirical engineering result: RSIC improves fidelity in user-specified regions by encoding extra local features and guiding the diffusion decoder. The derivation in Section II.C is a modeling approximation: Eq. (3) replaces the intractable conditioning term with a distortion gradient, which is a design choice rather than a quantity fitted to the evaluation metrics. The local-fidelity gain is not a prediction that reduces to a fit: the referring mode allocates more bits to the user-specified region by construction, and the evaluation measures fidelity in that region, but the paper reports measured performance rather than presenting the gain as a theoretical prediction derived from its own inputs. The only self-citation is [15] for the conditional prior in HSVLC; this is a component used as an off-the-shelf prior, not the claim being derived, and it is externally published work. The ablation 'w/o RGE' does not isolate γ=0 with identical bit allocation, but that is a missing control or confound, not circularity: no equation or fitted parameter is reused as evidence for the same equation. Therefore no circular step is present.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on several domain assumptions about pretrained models (Stable Diffusion, GEM) and on an ad hoc modeling of the guidance likelihood. The hyperparameters gamma, omega, scale boundaries, and M quantization are hand-picked and not justified by ablations. No new physical entities are introduced.

free parameters (5)
  • referring scale gamma = gamma = 10^3 * sqrt(alpha_{t-1}/alpha_t)
    Hand-picked in Section III.A; controls the strength of the latent-guidance term. No sensitivity analysis provided.
  • global guidance scale omega = omega = 3 * (1 - M * 0.7)
    Hand-designed in Section II.C to spatially vary the classifier-free guidance; no ablation.
  • representation scale boundaries = 1/2, 3/4, 7/8
    Empirically set in Section II.B, controlling which representation scale serves each M value. No ablation or justification.
  • M quantization levels = 1 to 8 levels
    Chosen to trade off bitrate of the weight map (max 0.0007 bpp) against fidelity; no ablation.
  • training lambda schedule = lambda = 0.01 * exp(7M)
    Follows reference [14], but no sensitivity analysis for this setting.
assumptions (4)
  • domain assumption Stable Diffusion's pretrained latent space is a valid representation for both lossy compression and generative reconstruction.
    The entire method operates on Stable Diffusion latents (z0) and relies on the autoencoder E/D and U-Net [10] for reconstruction; the paper does not validate this assumption, but it is consistent with prior work [8].
  • domain assumption DDIM inversion can recover noisy latents z_hat_t from z_hat_0 in the limit of small steps.
    Invoked in Section II.C, 'Based on the assumption that the denoising process can be reversed in the limit of small steps [18]', and used in Algorithm 1 to initialize the guidance loop.
  • ad hoc to paper The conditional likelihood p(z_hat_t | z_tilde_t) is proportional to exp(-||C(z_tilde_t) - z_hat_t||^2).
    In Section II.C, the paper jumps from Eq. (3)'s score decomposition to minimizing the distortion directly; this exponential-distortion model is not derived or validated.
  • domain assumption The grounding model GEM (reference [4]) can reliably localize user-specified content in arbitrary images.
    The weight map M is generated by GEM from a referring word; failures of GEM would directly misallocate bits, but no analysis of GEM's errors is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Your Demands Deserve More Bits: Referring Semantic Image Compression at Ultra-low Bitrate." pith.science (2026). https://pith.science/paper/H5WAYJJR

@misc{pith2026250600526,
  author       = {Pith},
  title        = {Pith review of: Your Demands Deserve More Bits: Referring Semantic Image Compression at Ultra-low Bitrate},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H5WAYJJR}},
  note         = {Machine review of arXiv:2506.00526}
}
read the original abstract

With the help of powerful generative models, Semantic Image Compression (SIC) has achieved impressive performance at ultra-low bitrate. However, due to coarse-grained visual-semantic alignment and inherent randomness, the reliability of SIC is seriously concerned for reconstructing completely different object instances, even they are semantically consistent with original images. To tackle this issue, we propose a novel Referring Semantic Image Compression (RSIC) framework to improve the fidelity of user-specified content while retaining extreme compression ratios. Specifically, RSIC consists of three modules: Global Description Encoding (GDE), Referring Guidance Encoding (RGE), and Guided Generative Decoding (GGD). GDE and RGE encode global semantic information and local features, respectively, while GGD handles the non-uniformly guided generative process based on the encoded information. In this way, our RSIC achieves flexible customized compression according to user demands, which better balance the local fidelity, global realism, semantic alignment, and bit overhead. Extensive experiments on three datasets verify the compression efficiency and flexibility of the proposed method.

Figures

Figures reproduced from arXiv: 2506.00526 by the authors.

Figure 1
Figure 1. The proposed RSIC can improve the fidelity of user-specified content [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Structure of the proposed Hierarchical Spatially Variable-rate Latent [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Guided generative decoding process. During generative decoding, RSIC adjusts the weights of two guidance to maintain a balance between fidelity and semantic consistency based on user demands. Specifically, RSIC employs a spatially non-uniform guidance scale defined as ω = 3×(1−M ∗0.7). This approach ensures low global se￾mantic guidance in regions with high M values. Additionally, regions with larger M values retain… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Evaluation of codecs on images from MS-COCO subset. [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison of the proposed method with other compression codecs, test images are selected from Kodak, CLIC, and MS-COCO datasets. [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 17 canonical work pages

  1. [15]

    High efficiency deep image compression via channel-wise scale adaptive latent representation learning,

    C. Wu, Q. Wu, K. N. Ngan, H. Li, F. Meng, and L. Xu, “High efficiency deep image compression via channel-wise scale adaptive latent representation learning,” Signal Processing: Image Communication, vol. 130, p. 117227, 2025

    work page 2025

  2. [1]

    The jpeg still picture compression standard,

    G. K. Wallace, “The jpeg still picture compression standard,” IEEE Transactions on Consumer Electronics , vol. 38, no. 1, pp. xviii–xxxiv, 1992

    work page 1992

  3. [2]

    Overview of the high efficiency video coding (hevc) standard,

    G. J. Sullivan, J.-R. Ohm, W.-J. Han, and T. Wiegand, “Overview of the high efficiency video coding (hevc) standard,” IEEE Transactions on Circuits and Systems for Video Technology , vol. 22, no. 12, pp. 1649– 1668, 2012

    work page 2012

  4. [3]

    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

    work page 2021

  5. [4]

    Grounding everything: Emerging localization properties in vision-language transformers,

    V . F. H. K. Walid Bousselham, Felix Petersen, “Grounding everything: Emerging localization properties in vision-language transformers,” arXiv preprint arXiv:2312.00878, 2023

    arXiv 2023

  6. [5]

    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,” in Proceedings of the 29th ACM international conference on multimedia , 2021, pp. 4230–4238

    work page 2021

  7. [6]

    Cross modal com- pression with variable rate prompt,

    J. Gao, J. Li, C. Jia, S. Wang, S. Ma, and W. Gao, “Cross modal com- pression with variable rate prompt,” IEEE Transactions on Multimedia , 2023

    work page 2023

  8. [7]

    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,” in ICML 2023 Workshop on Neural Compression: From Information Theory to Applications , 2023

    work page 2023

Show all 33 references
  1. [8]

    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,” in The Twelfth International Conference on Learning Representations , 2024. [Online]. Available: https://openreview.net/forum?id=ktdETU9JBg

    2024

  2. [9]

    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,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 3836–3847

    2023

  3. [10]

    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,” 2021

    2021

  4. [11]

    Visual instruction tuning,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” in NeurIPS, 2023

    2023

  5. [12]

    Improved baselines with visual instruction tuning,

    H. Liu, C. Li, Y . Li, and Y . J. Lee, “Improved baselines with visual instruction tuning,” 2023

    2023

  6. [13]

    zlib home site,

    “zlib home site,” https://www.zlib.net/

  7. [14]

    Variable-rate deep image compression through spatially-adaptive feature transform,

    M. Song, J. Choi, and B. Han, “Variable-rate deep image compression through spatially-adaptive feature transform,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 2380–2389

    2021

  8. [16]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,” in NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applica- tions, 2021

    2021

  9. [17]

    Score-based generative modeling through stochastic differ- ential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differ- ential equations,” arXiv preprint arXiv:2011.13456 , 2020

    2011 arXiv

  10. [18]

    Null- text inversion for editing real images using guided diffusion models,

    R. Mokady, A. Hertz, K. Aberman, Y . Pritch, and D. Cohen-Or, “Null- text inversion for editing real images using guided diffusion models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 6038–6047

    2023

  11. [19]

    Repaint: Inpainting using denoising diffusion probabilistic models,

    A. Lugmayr, M. Danelljan, A. Romero, F. Yu, R. Timofte, and L. Van Gool, “Repaint: Inpainting using denoising diffusion probabilistic models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 11 461–11 471

    2022

  12. [20]

    Universal guidance for diffusion models,

    A. Bansal, H.-M. Chu, A. Schwarzschild, S. Sengupta, M. Goldblum, J. Geiping, and T. Goldstein, “Universal guidance for diffusion models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 843–852

    2023

  13. [21]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” arXiv preprint arXiv:2010.02502 , 2020

    2010 arXiv

  14. [22]

    The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale,

    A. Kuznetsova, H. Rom, N. Alldrin, J. Uijlings, I. Krasin, J. Pont-Tuset, S. Kamali, S. Popov, M. Malloci, A. Kolesnikovet al., “The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale,” International journal of co...

    1956

  15. [23]

    True color kodak images,

    R. Franzen, “True color kodak images,” https://r0k.us/graphics/kodak/, November 1999

    1999

  16. [24]

    Workshop and challenge on learned image compression 2020,

    “Workshop and challenge on learned image compression 2020,” http: //compression.cc/

    2020

  17. [25]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick, “Microsoft coco: Common objects in context,” in Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13 . Springer,...

    2014

  18. [26]

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

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

    2017

  19. [27]

    Demysti- fying mmd gans,

    M. Bi ´nkowski, D. J. Sutherland, M. Arbel, and A. Gretton, “Demysti- fying mmd gans,” arXiv preprint arXiv:1801.01401 , 2018

    2018 arXiv

  20. [28]

    Clipscore: A reference-free evaluation metric for image captioning,

    J. Hessel, A. Holtzman, M. Forbes, R. L. Bras, and Y . Choi, “Clipscore: A reference-free evaluation metric for image captioning,” arXiv preprint arXiv:2104.08718, 2021

    2021 arXiv

  21. [29]

    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

  22. [30]

    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,” CoRR, vol. abs/2004.07728, 2020. [Online]. Available: https://arxiv.org/abs/2004. 07728

    2004 arXiv

  23. [31]

    Image harmonization with transformer,

    Z. Guo, D. Guo, H. Zheng, Z. Gu, B. Zheng, and J. Dong, “Image harmonization with transformer,” in Proceedings of the IEEE/CVF international conference on computer vision , 2021, pp. 14 870–14 879

    2021

  24. [32]

    Pct-net: Full res- olution image harmonization using pixel-wise color transformations,

    J. J. A. Guerreiro, M. Nakazawa, and B. Stenger, “Pct-net: Full res- olution image harmonization using pixel-wise color transformations,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 5917–5926

    2023

  25. [33]

    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,” in International Conference on Machine Learning. PMLR, 2023, pp. 25 426–25 443

    2023

Pith tools

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