SMIC: Semantic Multi-Item Compression based on CLIP dictionary

Thomas Maugey; Tom Bachard

REVIEW 4 major objections 5 minor 33 references

SMIC: Semantic Multi-Item Compression based on CLIP dictionary

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read CLIP's latent-space arithmetic is semantically linear, letting one shared dictionary compress entire image collections.

desk verdict A genuinely new multi-item semantic codec with a weak proof of its central linearity assumption; worth refereeing, but the semantic-fidelity headline is not supported. read the letter →

arxiv 2412.05035 v1 pith:5SCKVTS3 submitted 2024-12-06 eess.IV cs.CVcs.MM

Tom Bachard , Thomas Maugey This is my paper · ORCID

classification eess.IVcs.CVcs.MM

keywords semanticcompressionmulti-itemCLIPdictionarylearningsparsecodinggenerativerate-distortionimagecollections

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 tries to establish that CLIP's image-embedding space is semantically linear: adding and subtracting CLIP vectors adds and subtracts the high-level content of the corresponding images, and this property is enough to build a multi-image codec. The codec, called SMIC (Semantic Multi-Item Compression), learns a small semantic dictionary from a collection, encodes each image as a few sparse coefficients over that dictionary, and regenerates the image with a CLIP-conditioned diffusion generator. The authors report rates as low as around $10^{-5}$ bits per pixel per image (with a dictionary-inclusive rate of $1.4\times 10^{-4}$ BPP for their medium model on a 5000-image database), beating single-image CLIP codecs and classical video codecs while keeping comparable semantic fidelity. They also show that the learned atoms are interpretable high-level concepts and that the dictionary separates in-collection semantics from out-of-collection semantics. If the claim holds, collections of a few hundred images or more can be stored far more compactly than any per-image codec allows.

What carries the argument

The load-bearing object is the semantic linearity of CLIP's latent space, together with the learned semantic dictionary $\mathbf{T}$. Eq. (4) is the operation that lets latent additions and subtractions stand in for semantic additions and subtractions; Eq. (5) generalizes it to a weighted sum of dictionary atoms, so an image's semantics becomes a sparse coefficient vector over shared concepts. Eq. (8), an $\ell^1$-regularized projection solved by coordinate descent, recovers those coefficients for each image, and the quantized dictionary plus coefficient list is the entire compressed representation. The dictionary thus plays the role of the statistical side information in multi-item compression: it is transmitted once and its cost is amortized over the collection, while individual images remain randomly accessible.

What would settle it

Take an image whose content combines two concepts that rarely co-occur in natural images (e.g., a cathedral interior made of coral), compute its SMIC reconstruction with a dictionary trained on a disjoint set of scenes, and measure the semantic distance $d_\Phi(x,\hat{x})$ alongside the reconstruction error $\|z-\mathbf{T}c\|_2$; if a human annotator judges the regenerated image semantically unrelated while in-distribution images stay faithful, the linearity assumption is falsified for that input, and with it the universal rate-fidelity claim.

Watch

Extended reading notes

Core claim

The central claim is the semantic linearity identity $\Phi(\mathrm{CLIP}(x_1)+\lambda\,\mathrm{CLIP}(x_2))=\Phi(\mathrm{CLIP}(x_1))+\lambda\,\Phi(\mathrm{CLIP}(x_2))$, stated as Eq. (4): applying the CLIP-to-image generator $\Phi$ to a weighted sum of two CLIP embeddings is the same as generating each embedding separately and combining the resulting semantics. The authors extend this to multiple vectors, so any image whose concepts lie in the span of a learned dictionary $\mathbf{T}$ can be written as $z=\sum_j c_j t_j$, with the coefficients $c_j$ found by an $\ell^1$-penalized projection. The paper argues that the dictionary learned this way is genuinely semantic: each atom generates an interpretable scene (e.g., mountain, sea, sunset), and the sparse-coefficient representation preserves semantic coherence across dictionary sizes from $n_a=2$ to $n_a=128$. On this basis, SMIC transmits a quantized dictionary once and then, for each image, a short list of quantized coefficients; the decoder reconstructs the latent vector and generates the image with unCLIP. Against the single-image CLIP-unCLIP codec, SMIC's medium configuration achieves a rate of $1.4\times10^{-4}$ BPP, and the compression advantage begins at a collection size of 150 images.

Load-bearing premise

The method assumes that every image's meaning can be assembled by adding and subtracting a fixed set of dictionary concepts inside CLIP's embedding space; when two concepts were never seen together during CLIP's training, this addition stops working, and the codec has no detector to catch the failure.

Editorial extensions

If this is right

A collection of 150 images is already enough for SMIC to beat its single-image CLIP-unCLIP counterpart in total rate; for less competitive codecs the breakeven point is below 50 images.
The dictionary is learned once from the collection and can be reused for new images from the same distribution, so the collection can be added to without recompressing the existing items.
The sparse-coefficient representation is an interpretable semantic decomposition, so a user can edit or reconstruct an image by modifying individual coefficients or by adding or subtracting whole atoms.
Because the dictionary cost is fixed and the per-image rate is tiny, the rate advantage grows without bound as the collection size increases.
The projection/residual separation shows which semantics belong to the database and which do not, enabling a future semantic quantization scheme based on task-relevant concepts.

Reading between the lines

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

The linearity is only as reliable as CLIP's training distribution: for combinations of concepts never seen together, Eq. (4) fails, so a production codec would need an out-of-dictionary detector or a fallback single-image mode; the paper identifies the failure but does not implement detection.
A controlled test could map the failure boundary: take a set of concept pairs, measure where the reconstructed latent $\hat{z}$ stops generating the intended semantics, and check whether the semantic-fidelity gap correlates with the reconstruction error $\|z-\mathbf{T}c\|_2$; this would tell practitioners how far they can safely push the dictionary.
The same dictionary-as-semantic-filter idea suggests a task-oriented quantization: instead of spending bits uniformly, an agent could spend more coefficients on atoms whose semantics matter for the downstream task, which is the semantic quantization the authors leave as future work.
If other large multimodal encoders share this near-linear structure, the SMIC scheme or its dictionary-learning core could transfer to video collections or multimodal documents, where inter-item semantic redundancy is even larger.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The new thing here is real: a dictionary learned directly in CLIP latent space, used to compress a whole image collection with sparse coefficients, plus a nice demonstration that the learned atoms behave like semantic frequencies. That is a credible extension of the authors' own single-item CLIP codec and I have not seen it elsewhere. The random-access property is preserved, the dictionary overhead is analyzed honestly, and the paper openly concedes that latent addition fails when concepts were not seen together during training. That admission is more than many papers offer.

The soft spots are load-bearing. The "proof" of Eq. (4) is a handful of interpolated images, not a measurement, and the failure cases the authors admit in Section III are exactly the cases that matter for a codec: Eq. (8) discards the residual and then normalizes, so a projection that falls outside the additive span of the dictionary will produce a plausible but semantically wrong image. Fig. 10 already shows this for an out-of-collection input, and the paper does not check how often it happens for in-collection images. So the "not sacrificing semantic fidelity" claim in the abstract is not established.

The headline bitrate also deserves caution. Eq. (13) is a fitted approximation to the measured sparsity of Fig. 12, not a coded bitstream measurement, and the authors themselves call the exponent "debatable." The total rate in Eq. (14) inherits that fit. The parameter sets for the final comparisons are selected on the evaluation dataset itself, which softens the rate-distortion curves. Metrics CC, BSS, CSS are self-defined or drawn from prior work with no error bars. No code is released.

Having said that, the central idea is not invalidated by these issues. The dictionary projection is a reasonable coding scheme; the rate advantage is plausible even if the exact numbers are fuzzy; and the semantic separation experiment (projection vs. residual) is a genuinely interesting observation that could support future work. The comparison against single-item codecs is fine for the claim, since existing multi-item codecs work at pixel level and are not in the same regime.

Who gets value from this? Researchers in ultra-low-rate and machine-centric compression, and anyone pushing CLIP-based generative codecs. It deserves a serious referee, but the referee should push for code release, a real coded bitstream for the rate numbers, a quantitative test of linearity in-distribution, and a discussion of failure rates. I would not desk-reject it.

Referee Report

4 major / 5 minor

Summary. The paper proposes SMIC, a multi-item semantic compression scheme for image collections. Images are encoded with CLIP, a semantic dictionary is learned by sparse coding (Eq. 7), and each image's CLIP latent is approximated by a sparse combination of atoms; the coefficients and dictionary are quantized, and decoding uses Stable unCLIP to generate images. The authors claim that CLIP's latent space is semantically linear (Eq. 4), that the learned dictionary has a semantic interpretation, and that the resulting codec achieves extremely low rates (around 10^-5 BPP per image) without sacrificing semantic fidelity, outperforming single-item generative codecs.

Significance. The proposal is conceptually interesting and addresses a real gap: most semantic compression schemes are single-image, while collections have semantic redundancies. The paper provides a clear framework, includes a rate-distortion analysis with dictionary overhead, and honestly states limitations of the linearity assumption. If the linearity property holds sufficiently well and the rate model is accurate, the method could be valuable for cold storage and browsing of large collections. However, the central evidence is qualitative, the bitrate is computed from a fitted curve rather than measured, and the evaluation uses the same dataset for parameter selection and testing. These issues currently prevent the quantitative claims from being fully supported.

major comments (4)

[Section III, Eq. (4)] The paper claims to 'prove' semantic linearity, but the evidence is a handful of qualitative interpolation images (Figs. 4-5) and the text explicitly concedes that the addition fails when the added concepts were not seen together during CLIP training. Since every latent is approximated by a sparse sum of dictionary atoms (Eq. 8) and the residual is discarded, there is no mechanism to detect or correct images whose semantics are not in the additive span of the atoms. Figure 10 demonstrates exactly this failure for an out-of-collection image. The 'not sacrificing semantic fidelity' claim is therefore not established for general image collections; a quantitative evaluation of Eq. (4) over diverse concept pairs is needed.
[Section V-B, Eqs. (12)-(14)] The reported rates, including the headline comparisons in Tables III and IV, are computed using Eq. (13), which the authors describe in Section V-C as a 'rough' fit to Fig. 12 with a 'debatable' exponent. This is not a measurement of an actual coded bitstream. The rate claim should be validated by encoding the collection with the proposed coding scheme and measuring the total length, or by providing the fitted model with error bars and validating it on held-out data.
[Section V-C] The R-SF-O procedure selects the best parameter sets (na, λ, bdict, bcoef) on the Landscape dataset, and the same dataset is then used for the comparison in Table IV (n=5000). This selection on the evaluation set can inflate the reported rate-semantic performance. Please use a separate validation set for parameter selection and a held-out test set for the final comparison, or justify why overfitting is not a concern.
[Abstract vs Table IV] The abstract and contributions state a rate of 'around 10^-5 BPP per image', but Table IV reports 1.4×10^-4 BPP for the selected medium model at n=5000, and the n=∞ limit from Eq. (14) does not appear to reach 10^-5 for the given parameter sets. Please clarify the conditions for the 10^-5 figure and ensure the units and database size are consistent throughout.

minor comments (5)

[Section IV-C] The semantic metrics CC, BSS, and CSS are cited but not defined; please provide their formulas or precise equation references so the fidelity claims can be checked.
[Section IV-A, Eq. (5)] Eq. (5) has an index error: the sum runs over j but the atom is written as t_i; it should be t_j.
[Section V-B, Eq. (12)] The notation for the coefficient bit depth is inconsistent: Eq. (12) uses bc while the text and parameter lists use bcoef.
[Section V-C, Fig. 13] The Fig. 13 caption mentions n=1000 and n=∞, but the text says results are for n=100, 10000, 100000 and n=∞; please correct the caption.
[Abstract and Introduction] There are several grammatical errors in the abstract and introduction, e.g., 'shown to have attains better' and 'induce semantic additions and subtractions in the images are generated with UnCLIP'; a language edit would improve readability.

Circularity Check

1 steps flagged · score 6.0 of 10

Headline bitrate is an algebraic consequence of the authors' fitted sparsity law, not a measured bitstream; the rate claim reduces to its own fit.

fitted input called prediction [Section V-B, Eq. (13)-(14); used in Tables III and IV]
"From this figure, we derive the following equation for the proportion of non-null coefficients in the coefficients' list: P (c is null ) = 1 − P (c is non-null) ≈ 1/((λ + 1)^{log2(na)}) (13) Finally, the total compression rate of the compressed database, taking into account the dictionary overcost, is: Rtotal = na ∗ L ∗ bdict + N ∗ log2(na) ∗ bc / ((λ + 1)^{log2(na)}) (14)"

The reported per-image rate (Table IV: 1.4×10^-4 BPP) and the n* thresholds (Table III) are computed from Eq. (14), whose per-image term is an algebraic restatement of Eq. (13). Eq. (13) is not a measured bitstream length; it is a hand-fitted approximation to the authors' own measured sparsity in Fig. 12 ('where the exponent of (1+λ) is debatable'). Hence the headline compression gain is forced by the fitted P(null) curve: no entropy-coded stream is ever emitted or measured. This is the fitted-input-called-prediction pattern.

full rationale

The derivation chain for the central compression-rate claim passes through Eq. (14), which is exactly Eq. (13), a fit to the paper's own Fig. 12 sparsity measurements, multiplied by coding constants plus dictionary overhead. The paper never encodes an actual bitstream and never verifies the fitted formula against a real entropy-coded rate, so the reported 'around 10^-5 BPP' outperformance is a prediction built on its own fitted input. The other candidate circularities do not hold up: Eq. (4) is asserted and illustrated rather than derived, so its conceded failure cases are a soundness gap, not a circular step; the BSS/CSS metrics and the [21] baseline are prior self-cited published measurements, not definitions of the current output. Therefore the only exhibitable circularity is in the rate computation, making the central rate claim partially circular while the semantic-dictionary and fidelity experiments retain independent content.

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

The central claim rests on four tuned or fitted free parameters (na, lambda, bdict, bcoef, plus the fitted sparsity formula), on the asserted semantic linearity of CLIP, and on the authors' own evaluation metrics. No genuinely new entities are introduced; the dictionary is a learned matrix, not a postulated object. The ledger shows the paper contributes a new combination of known tools rather than a derivation from first principles.

free parameters (4)

Dictionary size na = 2 to 128; chosen models use 21 and 27
Number of atoms in T, selected by grid search in the rate-semantic-fidelity optimization (Section V-C); directly sets dictionary overhead na*L*bdict.
Sparsity weight lambda = 0.1, 0.2, 0.5, 1, 1.6, 2
Regularization weight in the L1 objectives (Eq. 7 and Eq. 8); controls the trade-off between reconstruction error and coefficient sparsity, and is tuned by hand.
Quantization bit depths bdict and bcoef = 16 default; 21 to 24 in the low/medium/high models
Uniform quantization levels for dictionary atoms and coefficients; selected during the convex-hull optimization and used to compute rates.
Sparsity rate model P(c null) = 1/((lambda+1)^log2(na)), exponent debatable
Eq. (13) is fitted to the measured null-coefficient proportions in Fig. 12 and then used to compute the total rate in Eq. (14); the authors note the exponent is debatable.

assumptions (5)

domain assumption CLIP latent space is semantically linear for natural concept combinations (Eq. 4).
Invoked in Section III and used to justify the dictionary decomposition in Section IV-A (Eq. 5); the paper demonstrates it only visually and admits failure for unseen concept combinations.
domain assumption Latents normalized to norm about 20 generate faithful images with unCLIP.
Inherited from prior work [21], used in Sections III and IV-D to normalize reconstructed latents before generation, but not re-verified here.
domain assumption L1 dictionary learning over CLIP latents produces semantically interpretable atoms.
Supported by qualitative atom images (Fig. 7) and a single decomposition example (Fig. 9); the semantic nature of atoms is asserted from these observations.
domain assumption CC, BSS and CSS measure semantic fidelity correctly.
All semantic fidelity conclusions use these self-defined, self-cited metrics from [21]; no independent or human evaluation is provided.
standard math L1 relaxation of the L0 sparse-coding objective preserves reconstruction quality.
Standard practice in dictionary learning, but the approximation gap is not quantified in this paper; used in Eq. (7).

how reviews work

0 comments

Cite this review

Pith. "Pith review of SMIC: Semantic Multi-Item Compression based on CLIP dictionary." pith.science (2026). https://pith.science/paper/5SCKVTS3

@misc{pith2026241205035,
  author       = {Pith},
  title        = {Pith review of: SMIC: Semantic Multi-Item Compression based on CLIP dictionary},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5SCKVTS3}},
  note         = {Machine review of arXiv:2412.05035}
}

abstract

Semantic compression, a compression scheme where the distortion metric, typically MSE, is replaced with semantic fidelity metrics, tends to become more and more popular. Most recent semantic compression schemes rely on the foundation model CLIP. In this work, we extend such a scheme to image collection compression, where inter-item redundancy is taken into account during the coding phase. For that purpose, we first show that CLIP's latent space allows for easy semantic additions and subtractions. From this property, we define a dictionary-based multi-item codec that outperforms state-of-the-art generative codec in terms of compression rate, around $10^{-5}$ BPP per image, while not sacrificing semantic fidelity. We also show that the learned dictionary is of a semantic nature and works as a semantic projector for the semantic content of images.

Figures

Figures reproduced from arXiv: 2412.05035 by the authors.

**Figure 1.** Multi-item compression with individual image coding. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 3.** Semantic Multi-item compression. IX describes the database’s statistics used for individual encoding and decoding. For X , a given collection of N images, SMIC framework aims at minimizing the following problem, where τΦ and 1https://huggingface.co/docs/diffusers/api/pipelines/stable unclip τΨ are respectively the semantic coherence threshold and the realism threshold: minX N i=1 R(zi) + R(IX ) s.t. (3) ∀i ∈ J1, NK… view at source ↗

Figures from the paper (11 more)

**Figure 4.** Figure 4: Progressively adding people to the landscape from [25] (Left) Input images [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Progressively removing the river from the landscape from [25]. (Left) Input images [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

**Figure 6.** Figure 6: Learning the dictionary from an image collection. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

**Figure 7.** Figure 7: Images generated from the first ten atoms of a dictionary learned of the whole landscape [27] dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

**Figure 8.** Figure 8: Generated images from their dictionary projection. The dictionary is learned on [27]. (Left to right) Input images. [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]

**Figure 9.** Figure 9: Example of decomposition in learned dictionary with [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]

**Figure 10.** Figure 10: Image generated from the projection and the residual of an image from [25] with a dictionary learned on [27]. (Left) [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]

**Figure 11.** Figure 11: Individual MIGC compression pipeline. To compress a database, this needs to be done on every image. [PITH_FULL_IMAGE:figures/full_fig_p009_11.png]

**Figure 12.** Figure 12: From this experiment, we observe that the proportion [PITH_FULL_IMAGE:figures/full_fig_p009_12.png]

**Figure 13.** Figure 13: Rate-distortion optimization curves for CC and the associated [PITH_FULL_IMAGE:figures/full_fig_p010_13.png]

**Figure 14.** Figure 14: Images generated from our different models. (Left to right) Input image. Image generated via the [PITH_FULL_IMAGE:figures/full_fig_p011_14.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 23 canonical work pages

[1]

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
[2]

Performance of the h. 263 video compression standard,

B. Girod, E. Steinbach, and N. F ¨arber, “Performance of the h. 263 video compression standard,” Journal of VLSI signal processing systems for signal, image and video technology , vol. 17, pp. 101–111, 1997

work page 1997
[3]

I. E. Richardson, The H. 264 advanced video compression standard . John Wiley & Sons, 2011

work page 2011
[4]

Compar- ative rate-distortion-complexity analysis of hevc and avc video codecs,

J. Vanne, M. Viitanen, T. D. Hamalainen, and A. Hallapuro, “Compar- ative rate-distortion-complexity analysis of hevc and avc video codecs,” IEEE Transactions on Circuits and Systems for Video Technology , vol. 22, no. 12, pp. 1885–1898, 2012

work page 2012
[5]

https://www.domo.com/learn/infographic/data-never-sleeps-11

“https://www.domo.com/learn/infographic/data-never-sleeps-11.” [Online]. Available: https://www.domo.com/learn/infographic/ data-never-sleeps-9

work page
[7]

Collective compression of images using averaging and transform coding,

P. M. Latha and A. A. Fathima, “Collective compression of images using averaging and transform coding,” Measurement, vol. 135, pp. 795–805, 2019

work page 2019
[8]

Novel image set compression algorithm using rate-distortion optimized multiple reference image selection,

L. Sha, W. Wu, and B. Li, “Novel image set compression algorithm using rate-distortion optimized multiple reference image selection,” IEEE Access, vol. 6, pp. 66 903–66 913, 2018

work page 2018
[9]

Learned Lossless Compression for JPEG via Frequency-Domain Prediction

J. Luo, S. Li, W. Dai, C. Li, J. Zou, and H. Xiong, “Learned lossless compression for jpeg via frequency-domain prediction,” 2023. [Online]. Available: https://arxiv.org/abs/2303.02666

work page Pith review arXiv 2023

Show all 33 references

[10]

Low-complexity and high-coding-efficiency image deletion for compressed image sets in cloud servers,

L. Sha, W. Wu, and B. Li, “Low-complexity and high-coding-efficiency image deletion for compressed image sets in cloud servers,” IEEE Transactions on Cloud Computing , vol. 11, no. 1, pp. 608–619, 2021

2021
[11]

Image set compression for similar images with priorities,

——, “Image set compression for similar images with priorities,” Electronics Letters , vol. 55, no. 5, pp. 262–264, 2019. [Online]. Available: https://ietresearch.onlinelibrary.wiley.com/doi/abs/10.1049/el. 2018.7342

2019
[12]

Lossless compression of jpeg coded photo collections,

H. Wu, X. Sun, J. Yang, W. Zeng, and F. Wu, “Lossless compression of jpeg coded photo collections,” IEEE Transactions on Image Processing, vol. 25, no. 6, pp. 2684–2696, 2016

2016
[13]

Semantic alignment for multi- item compression,

T. Bachard, A. J. Tom, and T. Maugey, “Semantic alignment for multi- item compression,” in 2022 IEEE International Conference on Image Processing (ICIP), 2022, pp. 2841–2845

2022
[14]

The perception-distortion tradeoff,

Y . Blau and T. Michaeli, “The perception-distortion tradeoff,” CoRR, vol. abs/1711.06077, 2017. [Online]. Available: http://arxiv.org/abs/ 1711.06077

2017 arXiv
[15]

Decision-making in cold chain logistics using data analytics: a literature review,

A. Chaudhuri, I. Dukovska-Popovska, N. Subramanian, H. K. Chan, and R. Bai, “Decision-making in cold chain logistics using data analytics: a literature review,” The International Journal of Logistics Management , vol. 29, no. 3, pp. 839–861, 2018

2018
[16]

Image coding for machines: an end-to-end learned approach,

N. Le, H. Zhang, F. Cricri, R. Ghaznavi-Youvalari, and E. Rahtu, “Image coding for machines: an end-to-end learned approach,” in ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 1590–1594

2021
[17]

Semantic compression with large language models,

H. Gilbert, M. Sandborn, D. C. Schmidt, J. Spencer-Smith, and J. White, “Semantic compression with large language models,” 2023. [Online]. Available: https://arxiv.org/abs/2304.12512

2023 arXiv
[18]

Semantic compression embedding for generative zero-shot learning

Z. Hong, S. Chen, G.-S. Xie, W. Yang, J. Zhao, Y . Shao, Q. Peng, and X. You, “Semantic compression embedding for generative zero-shot learning.” in IJCAI, 2022, pp. 956–963

2022
[19]

Generative adversarial networks,

I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial networks,” Communications of the ACM , vol. 63, no. 11, pp. 139–144, 2020

2020
[20]

Denoising diffusion probabilistic models,

J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” 2020. [Online]. Available: https://arxiv.org/abs/2006.11239

2020 arXiv
[21]

Can image compression rely on clip?

T. Bachard and T. Maugey, “Can image compression rely on clip?” IEEE Access, 2024

2024
[22]

Learning transferable visual models from natural language supervision,

A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” CoRR, vol. abs/2103.00020, 2021. [Online]. Available: https://ar...

2021 arXiv
[23]

Hierarchical text-conditional image generation with clip latents,

A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, and M. Chen, “Hierarchical text-conditional image generation with clip latents,” 2022

2022
[24]

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,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), 6 2022, pp. 10 684–10 695

2022
[25]

Kodak lossless true color image suite,

Kodakt, “Kodak lossless true color image suite,” 1999. [Online]. Available: https://r0k.us/graphics/kodak/

1999
[26]

Online dictionary learning for sparse coding,

J. Mairal, F. Bach, J. Ponce, and G. Sapiro, “Online dictionary learning for sparse coding,” in Proceedings of the 26th annual international conference on machine learning , 2009, pp. 689–696

2009
[27]

Histogan: Controlling colors of gan-generated and real images via color histograms,

M. Afifi, M. A. Brubaker, and M. S. Brown, “Histogan: Controlling colors of gan-generated and real images via color histograms,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021

2021
[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
[29]

Coordinate descent algorithms for lasso penalized regression,

T. T. Wu and K. Lange, “Coordinate descent algorithms for lasso penalized regression,” The Annals of Applied Statistics , vol. 2, no. 1,
[30]

Semantic compression with side information: A rate-distortion perspective,

T. Guo, Y . Wang, J. Han, H. Wu, B. Bai, and W. Han, “Semantic compression with side information: A rate-distortion perspective,” 2022. [Online]. Available: https://arxiv.org/abs/2208.06094

2022 arXiv
[31]

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

2023 arXiv
[32]

Devel- opments in international video coding standardization after avc, with an overview of versatile video coding (vvc),

B. Bross, J. Chen, J.-R. Ohm, G. J. Sullivan, and Y .-K. Wang, “Devel- opments in international video coding standardization after avc, with an overview of versatile video coding (vvc),” Proceedings of the IEEE, vol. 109, no. 9, pp. 1463–1493, 2021. Tom Bachard Tom Bachard gra...

2021
[2008]

Available: http://dx.doi.org/10.1214/07-AOAS147

[Online]. Available: http://dx.doi.org/10.1214/07-AOAS147

doi
[2021]

degree in Theoretical Computer Science from ´Ecole Normale Sup ´erieure de Rennes and Universit´e Rennes 1, Rennes, France, in 2021

He received an M.Sc. degree in Theoretical Computer Science from ´Ecole Normale Sup ´erieure de Rennes and Universit´e Rennes 1, Rennes, France, in 2021. He started his PhD under the supervision of Thomas Maugey in 2021, at INRIA Bretagne, Rennes, France. His research interest...

2021

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

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

[2] [2]

Performance of the h. 263 video compression standard,

B. Girod, E. Steinbach, and N. F ¨arber, “Performance of the h. 263 video compression standard,” Journal of VLSI signal processing systems for signal, image and video technology , vol. 17, pp. 101–111, 1997

work page 1997

[3] [3]

I. E. Richardson, The H. 264 advanced video compression standard . John Wiley & Sons, 2011

work page 2011

[4] [4]

Compar- ative rate-distortion-complexity analysis of hevc and avc video codecs,

J. Vanne, M. Viitanen, T. D. Hamalainen, and A. Hallapuro, “Compar- ative rate-distortion-complexity analysis of hevc and avc video codecs,” IEEE Transactions on Circuits and Systems for Video Technology , vol. 22, no. 12, pp. 1885–1898, 2012

work page 2012

[5] [5]

https://www.domo.com/learn/infographic/data-never-sleeps-11

“https://www.domo.com/learn/infographic/data-never-sleeps-11.” [Online]. Available: https://www.domo.com/learn/infographic/ data-never-sleeps-9

work page

[6] [7]

Collective compression of images using averaging and transform coding,

P. M. Latha and A. A. Fathima, “Collective compression of images using averaging and transform coding,” Measurement, vol. 135, pp. 795–805, 2019

work page 2019

[7] [8]

Novel image set compression algorithm using rate-distortion optimized multiple reference image selection,

L. Sha, W. Wu, and B. Li, “Novel image set compression algorithm using rate-distortion optimized multiple reference image selection,” IEEE Access, vol. 6, pp. 66 903–66 913, 2018

work page 2018

[8] [9]

Learned Lossless Compression for JPEG via Frequency-Domain Prediction

J. Luo, S. Li, W. Dai, C. Li, J. Zou, and H. Xiong, “Learned lossless compression for jpeg via frequency-domain prediction,” 2023. [Online]. Available: https://arxiv.org/abs/2303.02666

work page Pith review arXiv 2023