Structured Pruning and Quantization for Learned Image Compression

Fengqing Zhu; Md Adnan Faisal Hossain

REVIEW 4 major objections 4 minor 44 references

Structured Pruning and Quantization for Learned Image Compression

T0 review · 4 major / 4 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 Structured pruning guided by the rate-distortion loss can cut a learned image compression model's size by about 30 percent without rate-distortion loss, and pruning plus 8-bit quantization reaches about 6.6x smaller models.

desk verdict A useful but not yet conclusive engineering study on pruning and quantizing learned image compression models; the headline claim of no BD-Rate drop needs a better-controlled baseline comparison. read the letter →

arxiv 2506.01229 v1 pith:DULHQRWX submitted 2025-06-02 eess.IV

Md Adnan Faisal Hossain , Fengqing Zhu This is my paper

classification eess.IV

keywords learnedimagecompressionstructuredpruningfilterchannelneuralarchitecturesearchmodelquantizationrate-distortionBD-Rate

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 sets out to show that structured pruning—removing whole filters and then channels inside the remaining filters—can shrink a learned image compression model by about 30 percent without hurting rate-distortion performance, provided the layer-wise pruning ratios are chosen by a neural architecture search that uses the rate-distortion loss as its criterion. On the Kodak test set, the best configuration reports a BD-Rate of -2.00% at 29.8% parameter reduction, meaning the pruned model is slightly better than the uncompressed baseline. The paper further argues that pruning to a moderate sparsity and then quantizing weights and activations to 8 bits outperforms aggressive pruning alone at the same model size, reaching 6.6x compression with +8.12% BD-Rate. If these results hold, learned codecs could run on resource-limited hardware with a smaller memory footprint while keeping coding efficiency.

What carries the argument

The load-bearing machinery is the layer-wise pruning-ratio search driven by the rate-distortion loss. For each layer, filters (and then filter channels) are progressively removed in groups, the network is finetuned briefly, and the increase in the rate-distortion loss $\Delta L_{RD}$ is recorded; the chosen ratio is the largest one keeping $\Delta L_{RD}$ below a tolerance $\alpha$. An outer adaptive loop adjusts $\alpha$ until the total sparsity $S$ is within $\delta$ of a target $S_{target}$, so the search converts a model-size constraint into a per-layer pruning decision. The rate-distortion loss itself is the same objective used to train the codec, which is what lets pruning be optimized for compression performance rather than for classification accuracy.

What would settle it

Retrain the baseline with a longer schedule or verify from its rate-distortion curves that it has converged, then rerun the CHIP-based filters-plus-channels-plus-NAS pruning at a 30 percent sparsity target; if the pruned model no longer matches or beats the baseline's BD-Rate, the no-drop claim rests on an undertrained baseline rather than on the pruning method.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that the rate-distortion objective of a learned image codec can directly guide structured pruning. The proposed pipeline first prunes each convolutional layer along the output dimension using output feature maps, then prunes channels inside each surviving filter using input feature maps. A two-stage NAS chooses, for each layer, the largest pruning ratio whose induced increase in rate-distortion loss stays under a tolerance alpha, and an adaptive search adjusts alpha until the whole-model sparsity matches a target. The reported result is that at a 30 percent target sparsity, the pruned model beats the uncompressed model by 2.0 percent BD-Rate, and that layer-wise NAS-selected ratios outperform fixed ratios at every sparsity tested. For higher compression, the paper shows that pruning to 20-35 percent sparsity followed by 8-bit quantization gives higher model compression and better BD-Rate than pruning to 80 percent sparsity without quantization.

Load-bearing premise

The headline result of -2.0% BD-Rate at 30 percent sparsity assumes the uncompressed baseline model is fully trained, because the baseline receives only 90 epochs of training and the paper does not report its rate-distortion curves or training convergence.

Editorial extensions

If this is right

A roughly 30 percent parameter reduction with no rate-distortion loss is achievable for a mean-scale hyperprior learned image compression model when the pruning criterion is CHIP.
Pruning filters and filter channels together is consistently better than pruning filters alone under the same model-size constraint.
NAS-selected layer-wise pruning ratios beat fixed per-layer ratios, and the advantage widens as the target sparsity rises.
For larger model-size reductions, moderate sparsity plus 8-bit quantization dominates aggressive pruning alone: 6.6x model compression at +8.12% BD-Rate versus 4.6x at +29.35%.
The method provides a direct size-versus-performance trade-off knob through the target sparsity hyperparameter.

Reading between the lines

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

The absence of reported baseline rate-distortion curves leaves open the possibility that some apparent gains reflect recovery of an undertrained baseline; a direct check would be to compare pruned models against a baseline trained for twice as many epochs.
Because the NAS uses a 250-image calibration set drawn from the training distribution, the chosen layer-wise ratios may be sensitive to the calibration domain, and testing on images outside that distribution would show how well the ratios transfer.
The same rate-distortion-guided ratio search could be applied to other components of learned compression pipelines, such as the entropy model or autoregressive context, and to other LIC architectures, potentially compounding the size reduction.
The joint pruning-and-quantization result suggests quantization error is a cheaper way to spend the model-size budget than heavy pruning; a systematic sweep over sparsity and bit-width pairs could map the efficient frontier of model size versus rate-distortion performance.

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.

Desk Editor's Note

read the letter

I read this as a workshop-style paper, and it does something genuinely new: it applies filter and channel pruning together, with a NAS-selected layer-wise pruning ratio, to the full Mean Scale Hyperprior model, and then combines the pruned network with 8-bit quantization. Previous pruning work on LIC stopped at the hyperprior or the decoder. That application is worth having on record, and the empirical comparison across three pruning criteria (L2, HRANK, CHIP) is a reasonable way to show that feature-guided criteria transfer to lossy compression.

The strongest result is Table 1: CHIP + filter/channel pruning + NAS gives -2.00% BD-Rate at about 30% parameter reduction on Kodak. That is a good outcome and, if it holds up, makes the method practically attractive. I also like the message in Table 3 that moderate pruning plus quantization beats aggressive pruning at the same model size; that is a sensible and useful trade-off for deployment.

The soft spots are real, and they are the usual ones for this kind of paper. First, the results are single-run, with no error bars or seeds. A -2.00% BD-Rate could easily be a fluke or a lucky initialization. Second, the baseline is trained for only 90 epochs on COCO and there is no evidence it converged. The pruned models are finetuned during the NAS search and again for 60 epochs after the ratios are fixed, so they get substantially more gradient updates. The -2.00% may just be the pruned model recovering baseline performance through extra finetuning. Without baseline rate-distortion curves or a longer training schedule, the 'no BD-Rate drop' claim is not established. Third, the paper compares against its own uncompressed baseline but not against the cited quantization-only or decoder-pruning methods, so it is unclear whether the joint method is actually better than the existing alternatives.

I would not desk-reject this. The method is coherent, the paper is clearly written, and the quantization-plus-pruning result is a genuine practical data point. But I would ask the authors for repeated runs, baseline convergence evidence, and at least one comparison to a prior pruning-only LIC method. If the -2% survives that check, this becomes a solid contribution.

Referee Report

4 major / 4 minor

Summary. The paper proposes a structured pruning and quantization framework for learned image compression (LIC) models, specifically a Mean-Scale Hyperprior architecture. The method prunes both filters and filter channels, uses a two-stage Neural Architecture Search (NAS) to choose layer-wise pruning ratios based on the rate-distortion loss, and then combines pruning with 8-bit uniform quantization. Experiments on the Kodak dataset report BD-Rate relative to an uncompressed baseline and parameter reduction; the best configuration (CHIP criterion with filter+channel pruning and NAS) is reported to achieve BD-Rate of -2.00% at 29.8% parameter reduction. The source code is made publicly available.

Significance. If the reported results are reproducible and robust, the paper would make a practical contribution to model compression for LIC, a topic that is less explored than pruning for classification networks. The combination of filter and filter-channel pruning guided by a rate-distortion-aware NAS is a sensible idea, and the public code release is a concrete strength. However, the central claim of achieving model size reduction without BD-Rate degradation rests on a single configuration and on a comparison that is not controlled for optimization budget, so the significance cannot be fully assessed from the current evidence.

major comments (4)

[Sec. 4.1, Tables 1-3] All BD-Rate values in Tables 1-3 are single runs with no error bars, seeds, or significance tests. Since the headline claim of "no BD-Rate drop" depends on one configuration (-2.00% at 29.8% parameter reduction in Table 1), the reader cannot determine whether this is a real effect or within-run variation. Please report at least three independent runs with standard deviations and specify initialization and training details for each run.
[Sec. 4.1 vs Sec. 3.2 and Sec. 4.2] The comparison between the pruned model and the baseline is not controlled for optimization budget. The baseline is trained for 90 epochs on COCO (Sec. 4.1), while the pruned model receives additional finetuning on a 250-image calibration set repeatedly during the NAS search (Sec. 3.2) and then a further 60 epochs of finetuning (Sec. 4.2). If the 90-epoch baseline is not fully converged, the negative BD-Rate in Table 1 could reflect extra gradient updates rather than a benefit of pruning. Please include baseline training/validation curves and final rate-distortion curves, and compare against a baseline retrained for the same total number of gradient updates, or otherwise demonstrate that 90 epochs is sufficient for convergence.
[Sec. 4, Tables 1-3] The paper does not compare against existing quantization-only or decoder-pruning baselines for LIC, despite citing them in Sec. 1 (Refs. [8]-[15]). As a result, the claim that joint pruning and quantization is advantageous over prior LIC model-compression approaches is not experimentally supported. Please add at least one representative quantization baseline and one structured-pruning baseline evaluated under comparable model sizes and bit-widths.
[Eq. (2), Sec. 3.2] The NAS criterion assumes that the per-layer change in rate-distortion loss after brief finetuning, ∇L_RD, is a reliable monotone proxy for final BD-Rate after full finetuning, but this assumption is not validated. The hyperparameters α, S_target, δ, K, and the calibration-set size are fixed without sensitivity analysis, and no evidence shows that the selected layer-wise ratios are stable across initializations. Please provide a sensitivity study for α and K, and a sanity check that the ordering of ∇L_RD values correlates with the ordering of final BD-Rate values.

minor comments (4)

[Eq. (2)] The set ξ in Eq. (2) is written as {n : ∇L_RD^{(n)} < α}, but n should be restricted to the discrete grid of evaluated pruning steps; otherwise the maximum over n is not well defined. Please clarify that n ranges over the tested group sizes.
[Sec. 4.2] The arithmetic expressions such as "8.9%(7.17 + 1.73)%" are confusing because the plus sign denotes a difference between improvements; writing the expressions as |7.17 - (-1.73)| = 8.9% would be clearer.
[Fig. 3] In Fig. 3, the meaning of K is not stated in the caption; the reader has to infer from Sec. 3.2 that K is the group size for progressive pruning. Please define K in the caption or in the figure itself.
[Sec. 2.1 and Sec. 4.1] The paper describes the baseline as a "Mean Scale Hyperprior" model and cites Ref. [2], but Ref. [2] is the joint autoregressive and hierarchical priors model; the relation between the implemented architecture and the cited references should be clarified.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: pruning ratios are selected on a COCO calibration set via rate-distortion loss, while BD-Rate is measured on the held-out Kodak set.

full rationale

The paper's central claim is that structured pruning guided by a rate-distortion-loss NAS can reduce LIC model size without BD-Rate drop. Tracing the derivation chain, the NAS in Sec. 3.2 selects layer-wise pruning ratios by thresholding the change in rate-distortion loss on a 250-image COCO calibration set (Eq. 2), and the final reported metric is BD-Rate on the Kodak dataset (Sec. 4.1-4.2). The selection criterion and the evaluation metric are therefore not the same quantity on the same data, so the reported BD-Rate is not forced by construction. The baseline is an independently trained Mean-Scale Hyperprior from [2], and the pruning criteria (L2-Norm, HRANK, CHIP) come from external prior works; no self-citation is load-bearing for the pruning claim. The concern that the 90-epoch baseline may be undertrained relative to the pruned model's additional finetuning is a comparison-validity issue, not circularity: nothing in the paper's equations makes the BD-Rate result equivalent to the NAS search criterion. Accordingly, no specific circular step can be exhibited, and the appropriate finding is no significant circularity.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The method rests on standard pruning criteria from prior works, a representative calibration set, and the assumption that per-layer loss-change thresholds proxy final BD-Rate. No new entities are introduced. Alpha, Starget, delta, K, and bit-width are user-set rather than fitted to the test set.

free parameters (6)

alpha (loss tolerance) = adapted during search to match target sparsity
Eq. (2) and Sec. 3.2: threshold on change in rate-distortion loss; adjusted via adaptive search until |S - Starget| <= delta.
Starget (target sparsity) = 30%, 45%, 60%, 80% in experiments
Sec. 4.3 and Table 2: user-chosen model size constraint.
delta (sparsity tolerance) = 0.01
Sec. 4.3: termination threshold for the NAS search.
K (progressive pruning group size) = not specified
Sec. 3.2: group size for pruning filters and channels; affects search cost and precision.
calibration dataset size = 250 images from COCO training
Sec. 4.1: used by the NAS to measure the change in rate-distortion loss.
quantization bit-width = 8 bits
Sec. 4.4: uniform quantization to int8 for weights and activations.

assumptions (3)

standard math Pruning criteria (L2 norm, HRANK, CHIP) from cited literature are valid importance estimators for LIC models.
Sec. 4.1: all three are used without modification; no verification that their importance rankings transfer from classification CNNs to LIC.
domain assumption The calibration set of 250 COCO training images is representative of the Kodak evaluation distribution.
Sec. 4.1 and Sec. 3.2: NAS thresholds the change in rate-distortion loss on this set; if unrepresentative, selected pruning ratios may not transfer.
ad hoc to paper Per-layer change in rate-distortion loss after brief finetuning is a monotone proxy for final BD-Rate after full finetuning.
Sec. 3.2 Eq. (2): the method chooses the maximum pruned count with loss change below alpha, assuming small calibration-set loss increase predicts small Kodak BD-Rate increase.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Structured Pruning and Quantization for Learned Image Compression." pith.science (2026). https://pith.science/paper/DULHQRWX

@misc{pith2026250601229,
  author       = {Pith},
  title        = {Pith review of: Structured Pruning and Quantization for Learned Image Compression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DULHQRWX}},
  note         = {Machine review of arXiv:2506.01229}
}

read the original abstract

The high computational costs associated with large deep learning models significantly hinder their practical deployment. Model pruning has been widely explored in deep learning literature to reduce their computational burden, but its application has been largely limited to computer vision tasks such as image classification and object detection. In this work, we propose a structured pruning method targeted for Learned Image Compression (LIC) models that aims to reduce the computational costs associated with image compression while maintaining the rate-distortion performance. We employ a Neural Architecture Search (NAS) method based on the rate-distortion loss for computing the pruning ratio for each layer of the network. We compare our pruned model with the uncompressed LIC Model with same network architecture and show that it can achieve model size reduction without any BD-Rate performance drop. We further show that our pruning method can be integrated with model quantization to achieve further model compression while maintaining similar BD-Rate performance. We have made the source code available at gitlab.com/viper-purdue/lic-pruning.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 37 canonical work pages

[8]

Lossy im- age compression with quantized hierarchical vaes,

Zhihao Duan, Ming Lu, Zhan Ma, and Fengqing Zhu, “Lossy im- age compression with quantized hierarchical vaes,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2023, pp. 198–207

work page 2023
[15]

Learned image compression with fixed-point arithmetic,

Heming Sun, Lu Yu, and Jiro Katto, “Learned image compression with fixed-point arithmetic,” in 2021 Picture Coding Symposium (PCS) . IEEE, 2021, pp. 1–5

work page 2021
[1]

Structured Pruning and Quantization for Learned Image Compression

INTRODUCTION Learned Image Compression (LIC) [1, 2, 3, 4, 5] has recently shown immense progress surpassing the bounds of traditional codecs such as JPEG [6] and VTM [7]. However, LICs uti- lize deep learning models which have large memory require- ments and latency constraints, hindering their deployment on resource-limited hardware. To address these cha...

work page Pith review arXiv 2025
[2]

Learned Image Compression Learned Image Compression is a form of image source coding that operates in the transform coding paradigm

PRELIMINARIES 2.1. Learned Image Compression Learned Image Compression is a form of image source coding that operates in the transform coding paradigm. This involves transformation of the input image into a latent representation which is subsequently quantized and entropy-coded into bit- streams for transmission or storage using a prior distribution. In o...

work page
[3]

Next, we demon- strate our proposed NAS procedure for determining the layer- wise pruning ratio of the pruned model

METHOD In this section, we first introduce our method for simultane- ously pruning filters and filter channels. Next, we demon- strate our proposed NAS procedure for determining the layer- wise pruning ratio of the pruned model. Finally, we outline the joint pruning and quantization scheme. 3.1. Pruning Filters + Filter Channels Due to the superiority of ...

work page
[4]

Next, we show the coding performance of our proposed pruning methods across dif- ferent model sizes

EXPERIMENTS We compare the performance of our proposed pruning meth- ods using three different pruning criteria, L2-Norm [36], HRANK [18] and CHIP [20]. Next, we show the coding performance of our proposed pruning methods across dif- ferent model sizes. Finally, we discuss the performance gain achieved by performing quantization in conjunction with prunin...

work page
[5]

We perform experiments to show that simultane- ously pruning filters and filter channels can produce better Fig

CONCLUSION In this work, we propose a model compression scheme tar- geted for LIC models based on structured pruning and quan- tization. We perform experiments to show that simultane- ously pruning filters and filter channels can produce better Fig. 4: Comparison of the rate-distortion curve of our baseline LIC model with those of our pruned and quantized...

work page
[6]

Variational image compression with a scale hyper- prior,

Johannes Ball ´e, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston, “Variational image compression with a scale hyper- prior,” arXiv preprint arXiv:1802.01436, 2018

arXiv 2018

Show all 44 references

[7]

Joint autore- gressive and hierarchical priors for learned image compression,

David Minnen, Johannes Ball ´e, and George D Toderici, “Joint autore- gressive and hierarchical priors for learned image compression,” Ad- vances in neural information processing systems, vol. 31, 2018

2018
[9]

Qarv: Quantization-aware resnet vae for lossy image compression,

Zhihao Duan, Ming Lu, Jack Ma, Yuning Huang, Zhan Ma, and Fengqing Zhu, “Qarv: Quantization-aware resnet vae for lossy image compression,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

2023
[10]

Mlic: Multi-reference entropy model for learned image compression,

Wei Jiang, Jiayu Yang, Yongqi Zhai, Peirong Ning, Feng Gao, and Ronggang Wang, “Mlic: Multi-reference entropy model for learned image compression,” in Proceedings of the 31st ACM International Conference on Multimedia, 2023, pp. 7618–7627

2023
[11]

The jpeg still picture compression standard,

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

1991
[12]

Overview of the versatile video cod- ing (vvc) standard and its applications,

Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm, “Overview of the versatile video cod- ing (vvc) standard and its applications,” IEEE Transactions on Cir- cuits and Systems for Video Technology, vol. 31, no. 10, pp. 3736–3764, 2021

2021
[13]

Effi- cient neural image decoding via fixed-point inference,

Weixin Hong, Tong Chen, Ming Lu, Shiliang Pu, and Zhan Ma, “Effi- cient neural image decoding via fixed-point inference,” IEEE Transac- tions on Circuits and Systems for Video Technology, vol. 31, no. 9, pp. 3618–3630, 2020

2020
[14]

End-to-end learned image compression with fixed point weight quan- tization,

Heming Sun, Zhengxue Cheng, Masaru Takeuchi, and Jiro Katto, “End-to-end learned image compression with fixed point weight quan- tization,” in 2020 IEEE International Conference on Image Processing (ICIP). IEEE, 2020, pp. 3359–3363

2020
[16]

Rate-distortion optimized post- training quantization for learned image compression,

Junqi Shi, Ming Lu, and Zhan Ma, “Rate-distortion optimized post- training quantization for learned image compression,” IEEE Transac- tions on Circuits and Systems for Video Technology, 2023

2023
[17]

Integer quantized learned image compression,

Geun-Woo Jeon, SeungEun Yu, and Jong-Seok Lee, “Integer quantized learned image compression,” in 2023 IEEE International Conference on Image Processing (ICIP). IEEE, 2023, pp. 2755–2759

2023
[18]

Memory-efficient learned image compression with pruned hyperprior module,

Ao Luo, Heming Sun, Jinming Liu, and Jiro Katto, “Memory-efficient learned image compression with pruned hyperprior module,” in 2022 IEEE International Conference on Image Processing (ICIP) . IEEE, 2022, pp. 3061–3065

2022
[19]

Efficient decoder for learned image com- pression via structured pruning,

Liewen Liao, Shaohui Li, Jixiang Luo, Wenrui Dai, Chenglin Li, Junni Zou, and Hongkai Xiong, “Efficient decoder for learned image com- pression via structured pruning,” in 2022 Data Compression Confer- ence (DCC). IEEE, 2022, pp. 464–464

2022
[20]

Complexity scalable learning-based image decoding,

Tahsir Ahmed Munna and Jo ˜ao Ascenso, “Complexity scalable learning-based image decoding,” in 2023 IEEE International Confer- ence on Image Processing (ICIP). IEEE, 2023, pp. 1860–1864

2023
[21]

Unified data-free compression: Pruning and quantization without fine- tuning,

Shipeng Bai, Jun Chen, Xintian Shen, Yixuan Qian, and Yong Liu, “Unified data-free compression: Pruning and quantization without fine- tuning,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 5876–5885

2023
[22]

Filter prun- ing via geometric median for deep convolutional neural networks ac- celeration,

Yang He, Ping Liu, Ziwei Wang, Zhilan Hu, and Yi Yang, “Filter prun- ing via geometric median for deep convolutional neural networks ac- celeration,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 4340–4349

2019
[23]

Hrank: Filter pruning using high-rank feature map,

Mingbao Lin, Rongrong Ji, Yan Wang, Yichen Zhang, Baochang Zhang, Yonghong Tian, and Ling Shao, “Hrank: Filter pruning using high-rank feature map,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 1529–1538

2020
[24]

Scop: Scientific control for reliable neu- ral network pruning,

Yehui Tang, Yunhe Wang, Yixing Xu, Dacheng Tao, Chunjing Xu, Chao Xu, and Chang Xu, “Scop: Scientific control for reliable neu- ral network pruning,” Advances in Neural Information Processing Sys- tems, vol. 33, pp. 10936–10947, 2020

2020
[25]

Chip: Channel independence-based pruning for compact neural networks,

Yang Sui, Miao Yin, Yi Xie, Huy Phan, Saman Aliari Zonouz, and Bo Yuan, “Chip: Channel independence-based pruning for compact neural networks,” Advances in Neural Information Processing Systems, vol. 34, pp. 24604–24616, 2021

2021
[26]

Fast convnets using group-wise brain damage,

Vadim Lebedev and Victor Lempitsky, “Fast convnets using group-wise brain damage,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 2554–2564

2016
[27]

Pruning filter in filter,

Fanxu Meng, Hao Cheng, Ke Li, Huixiang Luo, Xiaowei Guo, Guang- ming Lu, and Xing Sun, “Pruning filter in filter,” Advances in Neural Information Processing Systems, vol. 33, pp. 17629–17640, 2020

2020
[28]

Pconv: The missing but desirable sparsity in dnn weight pruning for real-time execution on mobile de- vices,

Xiaolong Ma, Fu-Ming Guo, Wei Niu, Xue Lin, Jian Tang, Kaisheng Ma, Bin Ren, and Yanzhi Wang, “Pconv: The missing but desirable sparsity in dnn weight pruning for real-time execution on mobile de- vices,” in Proceedings of the AAAI conference on artificial intelligence, 2020, ...

2020
[29]

Learning both weights and connections for efficient neural network,

Song Han, Jeff Pool, John Tran, and William Dally, “Learning both weights and connections for efficient neural network,” Advances in neural information processing systems, vol. 28, 2015

2015
[30]

Importance estimation for neural network pruning,

Pavlo Molchanov, Arun Mallya, Stephen Tyree, Iuri Frosio, and Jan Kautz, “Importance estimation for neural network pruning,” in Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 11264–11272

2019
[31]

Soft filter pruning for accelerating deep convolutional neural net- works,

Yang He, Guoliang Kang, Xuanyi Dong, Yanwei Fu, and Yi Yang, “Soft filter pruning for accelerating deep convolutional neural net- works,” arXiv preprint arXiv:1808.06866, 2018

2018 arXiv
[32]

Amc: Automl for model compression and acceleration on mobile de- vices,

Yihui He, Ji Lin, Zhijian Liu, Hanrui Wang, Li-Jia Li, and Song Han, “Amc: Automl for model compression and acceleration on mobile de- vices,” in Proceedings of the European conference on computer vision (ECCV), 2018, pp. 784–800

2018
[33]

Metapruning: Meta learning for automatic neural network channel pruning,

Zechun Liu, Haoyuan Mu, Xiangyu Zhang, Zichao Guo, Xin Yang, Kwang-Ting Cheng, and Jian Sun, “Metapruning: Meta learning for automatic neural network channel pruning,” in Proceedings of the IEEE/CVF international conference on computer vision , 2019, pp. 3296–3305

2019
[34]

Dmcp: Differentiable markov channel pruning for neural networks,

Shaopeng Guo, Yujie Wang, Quanquan Li, and Junjie Yan, “Dmcp: Differentiable markov channel pruning for neural networks,” in Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 1539–1547

2020
[35]

Towards op- timal structured cnn pruning via generative adversarial learning,

Shaohui Lin, Rongrong Ji, Chenqian Yan, Baochang Zhang, Liujuan Cao, Qixiang Ye, Feiyue Huang, and David Doermann, “Towards op- timal structured cnn pruning via generative adversarial learning,” in Proceedings of the IEEE/CVF conference on computer vision and pat- tern recogni...

2019
[36]

Up or down? adaptive rounding for post- training quantization,

Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort, “Up or down? adaptive rounding for post- training quantization,” in International Conference on Machine Learn- ing. PMLR, 2020, pp. 7197–7206

2020
[37]

Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization,

Xiuying Wei, Ruihao Gong, Yuhang Li, Xianglong Liu, and Fengwei Yu, “Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization,” arXiv preprint arXiv:2203.05740, 2022

2022 arXiv
[38]

Pact: Parameterized clipping activation for quantized neural networks,

Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan, “Pact: Parameterized clipping activation for quantized neural networks,” arXiv preprint arXiv:1805.06085, 2018

2018 arXiv
[39]

Learned step size quantiza- tion,

Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha, “Learned step size quantiza- tion,” arXiv preprint arXiv:1902.08153, 2019

1902 arXiv
[40]

Network quan- tization with element-wise gradient scaling,

Junghyup Lee, Dohyung Kim, and Bumsub Ham, “Network quan- tization with element-wise gradient scaling,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 6448–6457

2021
[41]

Pruning filters for efficient convnets,

Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Pe- ter Graf, “Pruning filters for efficient convnets,” arXiv preprint arXiv:1608.08710, 2016

2016 arXiv
[42]

Mi- crosoft coco: Common objects in context,

Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll ´ar, and C Lawrence Zitnick, “Mi- crosoft coco: Common objects in context,” in Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6- 12, 2014, Proc...

2014
[43]

Calculation of average psnr differences between rd-curves,

Gisle Bjontegaard, “Calculation of average psnr differences between rd-curves,” ITU SG16 Doc. VCEG-M33, 2001

2001
[44]

Kodak lossless true color image suite (photocd pcd0992),

Eastman Kodak, “Kodak lossless true color image suite (photocd pcd0992),” URL http://r0k. us/graphics/kodak, vol. 6, 1993

1993

Pith tools

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

[1] [8]

Lossy im- age compression with quantized hierarchical vaes,

Zhihao Duan, Ming Lu, Zhan Ma, and Fengqing Zhu, “Lossy im- age compression with quantized hierarchical vaes,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2023, pp. 198–207

work page 2023

[2] [15]

Learned image compression with fixed-point arithmetic,

Heming Sun, Lu Yu, and Jiro Katto, “Learned image compression with fixed-point arithmetic,” in 2021 Picture Coding Symposium (PCS) . IEEE, 2021, pp. 1–5

work page 2021

[3] [1]

Structured Pruning and Quantization for Learned Image Compression

INTRODUCTION Learned Image Compression (LIC) [1, 2, 3, 4, 5] has recently shown immense progress surpassing the bounds of traditional codecs such as JPEG [6] and VTM [7]. However, LICs uti- lize deep learning models which have large memory require- ments and latency constraints, hindering their deployment on resource-limited hardware. To address these cha...

work page Pith review arXiv 2025

[4] [2]

Learned Image Compression Learned Image Compression is a form of image source coding that operates in the transform coding paradigm

PRELIMINARIES 2.1. Learned Image Compression Learned Image Compression is a form of image source coding that operates in the transform coding paradigm. This involves transformation of the input image into a latent representation which is subsequently quantized and entropy-coded into bit- streams for transmission or storage using a prior distribution. In o...

work page

[5] [3]

Next, we demon- strate our proposed NAS procedure for determining the layer- wise pruning ratio of the pruned model

METHOD In this section, we first introduce our method for simultane- ously pruning filters and filter channels. Next, we demon- strate our proposed NAS procedure for determining the layer- wise pruning ratio of the pruned model. Finally, we outline the joint pruning and quantization scheme. 3.1. Pruning Filters + Filter Channels Due to the superiority of ...

work page

[6] [4]

Next, we show the coding performance of our proposed pruning methods across dif- ferent model sizes

EXPERIMENTS We compare the performance of our proposed pruning meth- ods using three different pruning criteria, L2-Norm [36], HRANK [18] and CHIP [20]. Next, we show the coding performance of our proposed pruning methods across dif- ferent model sizes. Finally, we discuss the performance gain achieved by performing quantization in conjunction with prunin...

work page

[7] [5]

We perform experiments to show that simultane- ously pruning filters and filter channels can produce better Fig

CONCLUSION In this work, we propose a model compression scheme tar- geted for LIC models based on structured pruning and quan- tization. We perform experiments to show that simultane- ously pruning filters and filter channels can produce better Fig. 4: Comparison of the rate-distortion curve of our baseline LIC model with those of our pruned and quantized...

work page

[8] [6]

Variational image compression with a scale hyper- prior,

Johannes Ball ´e, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston, “Variational image compression with a scale hyper- prior,” arXiv preprint arXiv:1802.01436, 2018

arXiv 2018