N-O Cool-chic: reconcile fast encoding with lightweight decoding for neural image compression

Olivier D\'eforges; Pierrick Philippe; Th\'eo Ladune; Th\'eophile Blard; Xiaoran Jiang

REVIEW 3 major objections 5 minor 16 references

N-O Cool-chic: reconcile fast encoding with lightweight decoding for neural image compression

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read This paper introduces a non-overfitted variant of Cool-chic that trades 45% rate for 1000x faster encoding while preserving a 2300-MAC/pixel decoder.

desk verdict A useful engineering result—a fast encoder paired with a 2.3 kMAC/pixel decoder—but the headline 1000x encoding-complexity claim is against an unspecified Cool-chic baseline; the paper's own matched-quality comparison is 20x. read the letter →

arxiv 2607.13723 v1 pith:4G27ME6E submitted 2026-07-15 eess.IV

Théophile Blard , Théo Ladune , Pierrick Philippe , Xiaoran Jiang , Olivier Déforges This is my paper · ORCID

classification eess.IV

keywords neuralimagecompressionlow-complexitydecodingoverfittedcodecsCool-chicnon-overfittedencodinganalysistransformrate-distortionreal-time

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 aims to take Cool-chic, an overfitted neural codec with an extremely light decoder, and make its encoding practical by replacing per-image overfitting with a learned encoder network that produces the latent representation in a single forward pass. It claims this cuts encoding complexity by roughly a factor of 1000 while keeping decoding at 2300 MAC/pixel, at the cost of a 45% rate increase compared to fully overfitted Cool-chic. The result matters because it opens a middle path: neural codecs with autoencoder-like encoding speed but decoder costs low enough for smartphones and other power-constrained devices. The paper shows this variant stays within about 1% of a well-known 2018 learned codec and surpasses a 2023 model designed specifically for low-complexity decoding.

What carries the argument

The core mechanism is the analysis transform: a stack of ConvNeXt residual blocks with interleaved downsampling and 1x1 convolutions that maps an input image to seven hierarchical latent grids in a single forward pass. These latents are then quantized and fed into the existing Cool-chic decoder, which combines an auto-regressive probability model for entropy coding, an upsampling network, and a synthesis network. Because the decoder parameters are learned once on a training set and shared, the bitstream carries only the latents, and the expensive per-image overfitting loop is replaced by a single pass through the analysis network.

What would settle it

Run the original Cool-chic with early stopping on the same test set until it reaches the same rate-distortion point as N-O Cool-chic, then compare total encoder MACs or wall-clock time; if the ratio is close to 20 rather than 1000, the headline complexity claim fails. Alternatively, measure N-O Cool-chic's end-to-end encoding latency on a representative mobile CPU and compare it against the stated 160 kMAC/pixel figure.

Watch

Extended reading notes

Core claim

The paper's central claim is that overfitting is not required for Cool-chic's lightweight decoder to work well. By training an analysis transform on a large image dataset, the encoder can generate a multi-resolution latent representation compatible with the standard Cool-chic decoder in one forward pass, and the decoder parameters become shared across all images so they no longer need to be transmitted. Compared to the original overfitted Cool-chic, this non-overfitted version reduces encoding complexity from iterative gradient descent over latents and network weights to a single forward pass at roughly 160 kMAC/pixel, a factor of about 1000 lower. Decoding complexity remains at 2300 MAC/pix

Load-bearing premise

The claim of a 1000x encoding complexity reduction assumes the original Cool-chic is run to full overfitting convergence; the paper never specifies Cool-chic's total MAC count or training length, and when Cool-chic is trained only enough to match N-O Cool-chic's quality, the complexity gap shrinks to about 20x.

Editorial extensions

If this is right

Encoding an image becomes a single forward pass, roughly under one second, making Cool-chic-style coding viable for real-time and interactive applications.
The decoder remains extremely light at 2300 MAC/pixel, about 30 times lighter than the decoder of a 2018 learned codec and about 8 times lighter than a 2023 shallow-decoder model, while matching or beating their rate-distortion performance.
Shared decoder parameters eliminate the need to transmit per-image network weights, simplifying the bitstream and enabling a fixed decoder to be deployed once on client devices.
The 45% rate penalty relative to overfitted Cool-chic cleanly quantifies how much performance overfitting buys, providing a concrete trade-off curve between encoding complexity and compression efficiency.
With encoding complexity comparable to conventional autoencoder-based codecs, the approach makes ultra-light decoding available at nearly the same encoding cost as standard learned codecs.

Reading between the lines

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

The 1000x speedup figure depends on the baseline Cool-chic being trained to full convergence; the paper's own comparison at matching rate-distortion quality shows the gap is closer to 20x, still large but less dramatic. This suggests practitioners should compare at matched quality, not matched training budget.
A shared decoder opens the door to video and multi-image scenarios: the same decoder can be reused across frames, with only latent residuals transmitted, potentially making per-frame neural compression practical.
The encoder/decoder architecture is asymmetric by design, so one could deploy a fixed lightweight decoder on devices and periodically update a stronger encoder on a server, adapting to new content without changing the decoder.
The 45% gap to overfitted Cool-chic is a natural target for a hybrid scheme: a fast one-shot encoding plus a short per-image fine-tuning of just a few latent values could recover most of the lost performance at a fraction of full overfitting cost.

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

Quick take: this is a solid engineering paper with a real result—a single-forward-pass encoder (160 kMAC/pixel) feeding a 2.3 kMAC/pixel Cool-chic decoder, landing within about 1% of Ballé 2018 on CLIC 2020 validation. That's a genuinely useful operating point on the encoding/decoding complexity frontier. But the headline "1000x encoding complexity reduction" is not the right number to lead with. The paper itself shows why: at matched rate-distortion, the gap to Cool-chic is 20x, not 1000x. The abstract doesn't say that. It's a presentation problem, not a fraud, but it's the number people will remember.

What's new: attaching a learned analysis transform to the Cool-chic decoder, so you skip per-image overfitting entirely. The analysis is a stack of ConvNeXt residual blocks producing seven hierarchical latent grids, and the decoder stays the same lightweight one from their prior work. This is a natural combination, and I haven't seen it done for Cool-chic before. The measured trade-off is meaningful. The experiments are straightforward: RD curves on CLIC 2020 validation, a complexity table, and a clear comparison against Ballé, ELIC, and Yang 2023. To their credit, the 20x matched-quality number is in the body of the paper, so they're not hiding it.

Soft spots. First, the 1000x claim is computed against fully-converged Cool-chic, and the paper never states Cool-chic's total training MACs or iteration count. So the headline ratio is unauditable. The honest claim is "up to 20x at equal RD, up to 1000x if you insist on full convergence." Second, no code, no error bars, single validation set—standard for this subfield, but it means the ~1% gap to Ballé could easily be noise. Third, "competitive performance" in the abstract is vague; the body is more careful about saying it falls short of overfitted Cool-chic and state-of-the-art autoencoders.

Verdict: the central idea works, the paper is clearly written, and the authors are upfront about trade-offs in the body. The abstract overstates the gain. A referee should push them to lead with the matched-quality 20x figure and to specify the Cool-chic baseline precisely. I'd send it to review.

Referee Report

3 major / 5 minor

Summary. The paper proposes N-O Cool-chic, an extension of the overfitted Cool-chic codec that replaces per-image overfitting by a learned analysis transform (encoder network) which produces the latent representation in a single forward pass. The decoder is the same 2300 MAC/pixel Cool-chic decoder. The authors train the resulting autoencoder on CLIC 2019 patches and evaluate on CLIC 2020 validation. They report a 1000x reduction in encoding complexity relative to Cool-chic, a 45% rate increase relative to Cool-chic at equal quality, BD-rate results within 1% of Ballé 2018, and better BD-rate than Yang 2023. The paper concludes that the non-overfitted codec trades rate-distortion performance for much faster encoding while keeping the lightweight decoder.

Significance. The paper addresses a real bottleneck of overfitted codecs—encoding time—by pairing the lightweight Cool-chic decoder with a learned encoder. The proposed method is simple and the experimental comparisons cover relevant baselines. If the reported numbers hold, the contribution is useful for applications requiring fast encoding and very low decoder complexity. However, the headline 1000x gain is not yet well-supported because the Cool-chic baseline operating point is undefined, and the matched-quality gain is 20x per the paper's own Figure 1. The paper also lacks uncertainty quantification and code release, which limits confidence in the quantitative claims.

major comments (3)

[Abstract; Section 3, Figure 1] The headline claim of a 1000x encoding-complexity reduction is not well-defined. The paper never states Cool-chic's total encoding MAC count, number of iterations, or MACs per iteration. Figure 1 explicitly shows that Cool-chic's encoding complexity is varied by training time; Section 3 states that at a matched RD level Cool-chic requires 20x more MACs than N-O Cool-chic, not 1000x. Thus the 1000x figure compares against a fully-converged Cool-chic at a different (higher-quality) operating point, while the abstract's 'maintaining competitive performance' implies an equal-quality comparison. The abstract and Section 3 should report the matched-quality 20x figure as the primary claim, and fully specify the baseline used for 1000x (training length, MACs/iteration) for auditability.
[Section 4, Table 1] The claim 'Compared to Ballé [6], this gap is reduced to only 1%' is not supported by the reported BD-rate numbers. Table 1 gives Ballé 2018 BD-rate 12.9% vs. N-O Cool-chic 14.7%, a 1.8 percentage-point gap, or roughly 14% relative. If the 1% refers to a specific rate point (e.g., at 30 dB in Figure 4), that operating point should be identified; otherwise the statement is inconsistent with the table.
[Section 4, Table 1 and Figure 3] The N-O Cool-chic encoder complexity of 160 kMAC/pixel is reported without derivation. The analysis network in Figure 3 includes L=7 hierarchical levels, C=64 channels, ConvNeXt blocks with depthwise convolutions, and multiple residual blocks; the MAC count should be computed from these hyperparameters (or provided in a table) so the comparison with Cool-chic's training complexity is auditable. The paper should also state whether MAC counts for training include backward passes and how the 1000x ratio is obtained from MAC/pixel figures.

minor comments (5)

[Section 4, Figure 1] Typo: 'Rate-distorsion' should be 'Rate-distortion' in the Section 4 heading and Figure 1 axis.
[Figure 1 caption] The caption says 'The circle radius denotes the decoding complexity (see Table 1)' but no scale is given; without a numeric scale, the relative circle sizes are not interpretable.
[References] Reference [9] links to the CLIC 2021 tasks page, but the text refers to CLIC 2020. Please verify the correct dataset and URL.
[Section 5] The conclusion states that encoding takes 'less than 1 second', but no wall-clock measurements are reported in the experiments. Either add timing measurements or soften this claim.
[Section 4] The reported BD-rate and MAC numbers come from a single training run; no error bars or number of seeds are given. At minimum, state the number of random seeds, or report variance across validation images.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: N-O Cool-chic is an empirical autoencoder variant benchmarked against external codecs; no prediction reduces to a fitted input or self-cited derivation.

full rationale

The paper's derivation chain is: build an analysis transform f_alpha that outputs quantized latents in one forward pass (Eq. 2), train it jointly with the fixed Cool-chic decoder by minimizing E_x[D + lambda R] (Eq. 3), and measure RD and complexity on CLIC 2020 validation against external codecs (HEVC, VVC, ELIC, Ballé 2018, Yang 2023). Nothing in this chain is defined in terms of the target result. The decoder is taken from the authors' prior Cool-chic work (refs. 10-12), but it is reused as a fixed component, not invoked as a uniqueness theorem or as the justification of the new encoder's performance; the new result is the empirical RD table and Figure 4. The '1000x encoding complexity' claim compares against an unspecified Cool-chic training operating point, and the paper itself reports a 20x gap at matched RD (Section 3, Figure 1); this is an underspecified-baseline/correctness concern, not an equation-level circularity or a fitted parameter renamed as a prediction. No fitted values are used to reproduce the headline numbers, and the benchmarks are external. Therefore no circular step meeting the quoted-evidence standard is present.

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

The central claim rests on standard rate-distortion training and the transferability of the Cool-chic decoder to latents produced by a different encoder. No new entities are introduced. The main free parameters are architecture choices and rate points, none of which are fitted to test data.

free parameters (3)

number of latent levels L = 7
Chosen by hand; higher L increases decoder complexity; no ablation.
channel count C (per-level features) = 64
Chosen by hand; controls encoder capacity; no ablation.
Lagrange multipliers λ = {0.02, 0.004, 0.001, 0.0004, 0.0001}
Selected to span rate range; standard practice in image compression; not fitted to test data.

assumptions (4)

standard math Minimizing MSE distortion plus entropy-estimated rate approximates perceptual quality (standard rate-distortion optimization)
Used throughout; MSE is the distortion metric.
domain assumption The entropy model p_psi and range coder faithfully code the latents at the estimated rate
Assumed; no verification of entropy coding efficiency.
domain assumption CLIC 2019 training set is representative of natural images for generalization
Training on random crops from CLIC 2019; evaluation on CLIC 2020 val.
domain assumption The overfitted Cool-chic decoder (2300 MAC/pixel) remains effective when decoder parameters are shared across images
Core to the method; not automatically guaranteed, but empirically evaluated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of N-O Cool-chic: reconcile fast encoding with lightweight decoding for neural image compression." pith.science (2026). https://pith.science/paper/4G27ME6E

@misc{pith2026260713723,
  author       = {Pith},
  title        = {Pith review of: N-O Cool-chic: reconcile fast encoding with lightweight decoding for neural image compression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4G27ME6E}},
  note         = {Machine review of arXiv:2607.13723}
}

read the original abstract

Overfitted image codecs achieve strong compression performance and low decoder complexity by learning a lightweight decoder for each image. Such codecs include Cool-chic, which presents image coding performance on par with VVC while requiring around 2000 multiplications per decoded pixel. However, the encoding time associated with overfitted codecs may be prohibitively long for real-time applications, posing a challenge to their practical implementation in such scenarios. To address this issue, this paper proposes to decrease the encoding complexity of Cool-chic by bypassing the overfitting procedure and complementing the decoder with an encoder network. The proposed non-overfitted (N-O) Cool-chic, significantly reduces encoding complexity by a factor of 1000 compared to Cool-chic, while maintaining competitive performance.

Figures

Figures reproduced from arXiv: 2607.13723 by the authors.

**Figure 1.** Rate-distorsion performance as a function of the encoding complexity on CLIC 2020 validation set [9]. Negative results: less rate is required to get the same quality than HEVC. Cool-chic encoding complexity is varied by adjusting the training time. The circle radius denotes the decoding complexity (see [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Cool-chic decoding. ARM: Auto-Regressive Model. This paper aims to bridge the gap between autoencoderbased codecs and overfitted codecs by combining the existing Cool-chic lightweight decoder with an encoder which generates the latent representation in a single forward pass. The proposed network forms an autoencoder that offers the benefits of both low encoding and decoding complexity. 2 Background: Cool-chic This… view at source ↗

**Figure 3.** Proposed analysis transform. DWConv stands for depth-wise convolution and ↓ denotes the stride. Decoder. The decoder parameters (ψ, υ, θ) are learned alongside the analysis transform fα. The architecture from [12] with a complexity of 2300 MAC (multiplicationaccumulation) per decoded pixel is selected. It is worth noting that the network parameters no longer need to be transmitted alongside the latent representatio… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Rate-distortion performance on the CLIC 2020 validation set [9]. to N-O Cool-chic to reach the same level of performance. Beside the reduction in MAC, encoding images with N-O Cool-chic is conceptually simpler since there is no more optimization through gradient descen…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 3 linked inside Pith

[6]

Variational image compression with a scale hyperprior

Johannes Ballé, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. InIn- ternational Conference on Learning Representations, 2018

2018
[1]

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

Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. ELIC: Efficient learned image compression with unevenly grouped space- channel contextual adaptive coding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022

2022
[2]

MLIC++: Lin- ear complexity multi-reference entropy modeling for learned image compression

Wei Jiang and Ronggang Wang. MLIC++: Lin- ear complexity multi-reference entropy modeling for learned image compression. InICML 2023 Work- shop Neural Compression: From Information Theory to Applications

2023
[3]

Sullivan et al

Gary J. Sullivan et al. Overview of the high efficiency video coding (HEVC) standard.IEEE Transactions on Circuits and Systems for Video Technology, 2012

2012
[4]

Bross et al

B. Bross et al. Overview of the versatile video coding (VVC) standard and its applications.IEEE Transac- tions on Circuits and Systems for Video Technology, 2021

2021
[5]

Computationally efficient neural im- age compression.arXiv preprint arXiv:1912.08771, 2019

Nick Johnston, Elad Eban, Ariel Gordon, and Jo- hannes Ballé. Computationally efficient neural im- age compression.arXiv preprint arXiv:1912.08771, 2019

arXiv 1912
[7]

EVC: Towards real-time neural image compression with mask decay

Wang Guo-Hua, Jiahao Li, Bin Li, and Yan Lu. EVC: Towards real-time neural image compression with mask decay. InThe Eleventh International Confer- ence on Learning Representations, 2023

2023
[8]

Computationally- efficient neural image compression with shallow de- coders

Yibo Yang and Stephan Mandt. Computationally- efficient neural image compression with shallow de- coders. InIEEE/CVF International Conference on Computer Vision, ICCV 2023, pages 530–540. IEEE

2023

Show all 16 references

[9]

Challenge on learned image coding 2020

CLIC20. Challenge on learned image coding 2020. http://clic.compression.cc/2021/tasks/index.html, 2020

2020
[10]

COOL-CHIC: coordinate-based low complexity hierarchical image codec

Théo Ladune, Pierrick Philippe, Félix Henry, Gor- don Clare, and Thomas Leguay. COOL-CHIC: coordinate-based low complexity hierarchical image codec. InIEEE/CVF International Conference on Computer Vision, ICCV 2023

2023
[11]

Low-complexity overfitted neural image codec

Thomas Leguay, Théo Ladune, Pierrick Philippe, Gordon Clare, Félix Henry, and Olivier Déforges. Low-complexity overfitted neural image codec. In 25th IEEE International Workshop on Multimedia Signal Processing, MMSP 2023

2023
[12]

Overfitted image coding at reduced complexity.arXiv preprint arXiv:2403.11651, 2024

Théophile Blard, Théo Ladune, Pierrick Philippe, Gordon Clare, Xiaoran Jiang, and Olivier Déforges. Overfitted image coding at reduced complexity.arXiv preprint arXiv:2403.11651, 2024

2024 arXiv
[13]

C3: high-performance and low-complexity neural compression from a single image or video.CoRR, 2023

Hyunjik Kim, Matthias Bauer, Lucas Theis, Jonathan Richard Schwarz, and Emilien Dupont. C3: high-performance and low-complexity neural compression from a single image or video.CoRR, 2023

2023
[14]

A convnet for the 2020s

Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. InProceedings of the IEEE/CVF conference on computer vision and pat- tern recognition, pages 11976–11986, 2022

2022
[15]

Bag of tricks for image clas- sification with convolutional neural networks

Tong He, Zhi Zhang, Hang Zhang, Zhongyue Zhang, Junyuan Xie, and Mu Li. Bag of tricks for image clas- sification with convolutional neural networks. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 558–567, 2019

2019
[16]

Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

2014 arXiv

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [6]

Variational image compression with a scale hyperprior

Johannes Ballé, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. InIn- ternational Conference on Learning Representations, 2018

2018

[2] [1]

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

Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. ELIC: Efficient learned image compression with unevenly grouped space- channel contextual adaptive coding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022

2022

[3] [2]

MLIC++: Lin- ear complexity multi-reference entropy modeling for learned image compression

Wei Jiang and Ronggang Wang. MLIC++: Lin- ear complexity multi-reference entropy modeling for learned image compression. InICML 2023 Work- shop Neural Compression: From Information Theory to Applications

2023

[4] [3]

Sullivan et al

Gary J. Sullivan et al. Overview of the high efficiency video coding (HEVC) standard.IEEE Transactions on Circuits and Systems for Video Technology, 2012

2012

[5] [4]

Bross et al

B. Bross et al. Overview of the versatile video coding (VVC) standard and its applications.IEEE Transac- tions on Circuits and Systems for Video Technology, 2021

2021

[6] [5]

Computationally efficient neural im- age compression.arXiv preprint arXiv:1912.08771, 2019

Nick Johnston, Elad Eban, Ariel Gordon, and Jo- hannes Ballé. Computationally efficient neural im- age compression.arXiv preprint arXiv:1912.08771, 2019

arXiv 1912

[7] [7]

EVC: Towards real-time neural image compression with mask decay

Wang Guo-Hua, Jiahao Li, Bin Li, and Yan Lu. EVC: Towards real-time neural image compression with mask decay. InThe Eleventh International Confer- ence on Learning Representations, 2023

2023

[8] [8]

Computationally- efficient neural image compression with shallow de- coders

Yibo Yang and Stephan Mandt. Computationally- efficient neural image compression with shallow de- coders. InIEEE/CVF International Conference on Computer Vision, ICCV 2023, pages 530–540. IEEE

2023