acceptodds
Under review as a conference paper at ICLR 2027

Self-Contained Probability Representation for Deployable Lossless Image Compression

Abstract

Recent learning-based lossless image compression methods achieve strong compression, but they rely on large pre-trained models whose heavy inference cost hinders practical deployment. To this end, we introduce a compact representation of image probability which is embedded in a bitstream, shortly. The representation is compiled once into a lookup table, eliminating repeated neural inference during coding. For deployable on-device encoding, we marginalize image-specific features into a reusable table without per-image fitting, while a compact tree restores self-contained delivery. Our method achieves 95.5/52.8 faster decoding/encoding than the fastest evaluated learned baseline, and 4/5.7 faster decoding/encoding than classical codecs, while maintaining competitive rate–throughput trade-offs. Experiments demonstrated capabilities beyond prior learned codecs, including 16-bit image compression, practical encoding and decoding on devices ranging from servers to wearables, and a lossy extension through bit-plane truncation. Self-contained decoding demo files are included in the supplementary material.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Lossless image compression, learned image compression, entropy coding, probability modeling, implicit neural representations, lookup tables, on-device compression

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_5p7yOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper explores lightweight lossless image coding, where probability distributions are embedded in the bitstream as lookup tables or compact representations. It proposes three variants for different coding scenarios: IPR-I fits a small neural model with Haar-based spatial coding for each image and stores its parameters in the bitstream; IPR-L combines predictions from a reference-image model with empirical counts to build reusable probability tables; and IPR-L+MDL encodes the image-specific probability representation as a tree, removing the decoder's reliance on external tables. The paper presents a fairly complete system with cross-device evaluations, demonstrating an attractive trade-off between bitrate and throughput.

Critical strengths
  1. The paper clearly distinguishes three scenarios: server-side pre-encoding, repeated local encoding, and fully self-contained transmission. This makes the design trade-offs between the variants easy to understand.
  2. The experiments cover a wide range of practical platforms, including GPU servers, Jetson, Raspberry Pi, iPhone, Mac, and DGX, suggesting a strong potential on real-world deployment.
Critical weaknesses
  1. My primary concern is the method mainly combines established algorithms and techniques, including bit-plane decomposition, LUT, INR, and MDL trees. Similar compositions have appeared in prior work: JPEG 2000 involves bit-plane coding; FSAR uses neural networks to model finite-state probabilities and generate lookup tables; COOL-CHIC optimizes and transmits compact neural probability models for individual images; and FLIF/MANIAC uses image-adaptive context trees for probability modeling. Although these methods differ from the specific implementation proposed here, they make it important to clarify what is genuinely new. As it stands, the engineering contribution is clear, but the novelty beyond system integration remains insufficiently established.

  2. The learnable component mainly lies in the conditional probability representation, while the color transform, wavelet transform, bit-plane processing, and entropy-coding pipeline rely on fixed techniques. Thus, lossless reconstruction itself does not depend on neural networks. The value of the learnable component should instead be demonstrated clearly. However, the paper does not sufficiently compare the proposed neural probability models with such a statistical baseline. It is therefore difficult to determine how much of the reported performance comes from the learnable representation, as opposed to the existing transforms, context design, and coding implementation.

Minor Comment

  • In Line 161, the format of lookup table TbT_b appears inconsistent with Equation (4).
  • In Line 228, the citation of KT smoothing is not separated from the main text.
Questions for authors
  1. Beyond integrating existing techniques and algorithm, improving execution efficiency, and demonstrating cross-platform deployment, does the paper introduce a new algorithmic mechanism, network design, or analytical insight?

  2. How much does the learnable probability representation improve over a statistical baseline? With the reversible transforms, context definitions, and entropy-coding settings held constant, how does the neural probability table compare with one estimated from empirical frequencies and statistical smoothing, without neural training?

3 comments

Response to Reviewer 5p7y

Overall Response: We appreciate the reviewer for separating the engineering contribution from the role of the learned probability model. To illustrate the practical strengths highlighted in the review, we added Fig. 7, which shows IPR’s deployment scenarios. We also clarify the technical contribution relative to the cited work and report a controlled comparison with statistical probability estimation. All changes in the revised manuscript are highlighted in blue.


Q1/W1. Contribution relative to prior work

TL;DR IPR-I uses a finite spatial code to make an image-specific probability model compact to transmit and usable by lookup during lossless decoding.

Thank you for raising this point. As the reviewer notes, IPR builds on established coding techniques. Our design starts from a practical question:

Can a learned model be represented compactly enough to be transmitted with each file, yet remain enumerable so that lossless decoding requires only table lookups?

This imposes two requirements:

  1. The decoder should not need a shared pre-trained model.

  2. Entropy decoding should avoid per-symbol neural inference.

The table shows whether each codec requires a separately installed trained model and where it uses neural computation during decoding.

CodecNeural entropy modelSelf-containedNo Installed trained modelLosslessImage-specific adaptationNeural work at decoding
JPEG 2000 (EBCOT)NoYesYesYesAdaptive context probabilitiesNone
FLIF / MANIACNoYesYesYesImage-specific tree and adaptive probabilitiesNone
FSARYesNoNoYesShared trained modelLUT for latents; trained backbone required
COOL-CHICYesYesYesNoPer-image fittingLatent probability estimation and image synthesis
IPR-IYesYesYesYesPer-image fittingCompile LUT; decode by lookup

IPR-I meets both requirements for lossless decoding while retaining a learned, image-specific probability model. By quantizing the learned spatial features into a finite code, IPR-I lets the decoder reconstruct the probability table from the file and decode the transform bit-planes by lookup.

We added this comparison to the paper as Fig. 2(b).

Beyond this system design, the paper contributes the following analyses, which address the reviewer's question about new design or analytical insight:

Representation design & Analytical insights

  • Finite spatial conditioning (Table 4). A small MLP conditions on the causal state and a quantized spatial code built from 32 of 512 localized Haar atoms per patch, capturing spatial variation in bit probabilities while keeping the table finite. On Kodak, Haar: 3.1386 · NeRF-style: 3.1693 · Sinusoidal: 3.1939 · no spatial conditioning: 3.1795 bpsp.
  • Model cost vs. coding gain (Fig. 5). Relative to IPR-L's total rate, IPR-I adds 0.4–1.3% in model bits, saves 2.3–6.1% in payload, and reduces the total rate by 1.7–5.3%.
  • Learned prior (Appendix Fig. 10). Sparse states account for 72–81% of the log-loss gain over KT-only on held-out Kodak.

Q2/W2. Controlled comparison with statistical estimation

TL;DR At the same coding state, IPR-I reduces total file size by 8.56% on Kodak24 and 9.19% on DIV2K100 relative to Adaptive KT, which transmits no probability table.

Both baselines use the same state ss, symbol order, probability precision, and range coder as IPR-I without spatial conditioning. Both use the KT estimate p(1∣s)=(ns,1+1/2)/(ns,0+ns,1+1)p(1\mid s)=(n_{s,1}+1/2)/(n_{s,0}+n_{s,1}+1), but differ in how they obtain the counts:

  • Static KT: Counts come from the entire image. The encoder transmits a fixed table over occupied states, with sorted state indices delta-coded and table records compressed using zlib-9.
  • Adaptive KT: Counts start at zero for each independent 128×128128\times128 patch stream. The encoder and decoder update them identically after each bit. No table is transmitted.
MethodKodak24 total bpsp ↓DIV2K100 total bpsp ↓
IPR-I3.13863.0717
IPR-I (no spatial conditioning)3.17953.1525
Static KT (s)(s)4.49473.4898
Adaptive KT (s)(s)3.47713.4714

For the reusable IPR-L table,Appendix Table 10 and Fig. 10 compare count-only KT with the model-derived prior.

We added these rows to Table 4.


Minor comments

We appreciate the reviewer's careful reading. We made the lookup-table notation consistent between Eq. (4) and Line 161, and corrected the placement of the KT smoothing citation in Line 228.

Thank you for your response. I appreciate the additional results added to Table 4, which resolve my question relating to the statistical estimation. I tend to keep my positive assessment towards acceptance, particularly due to the comprehensive engineering contribution IPR has presented.

Thank you for your thoughtful feedback and support. We’re glad the additional results resolved your concern. We appreciate your recognition of our engineering contribution.

Reviewer_s1diOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work introduces a fast lossless image compression method IPR by introducing a compact representation of image probability which is embedded in a bitstream. Although the compression ratio is far from SOTA, the enc-dec throughput is faster than most current methods.

Critical strengths
  1. This method has a very fast decoding speed, and the article demonstrates the results of its deployment on different platforms.
Critical weaknesses
  1. This method's performance is still somewhat weaker compared to the state-of-the-art (SOTA) solution.
Questions for authors

Q1: As shown in figure 2, IPR-L achieves the fastest enc-dec throughput against other method. Is IPR-L the fastest method when compared with all other methods (including learning-base methods and traditional methods)?

Q2: is IPR a plug-in that can be directly applied to other existing methods? Can it improve the compression performance or efficiency?

Q3: IPR seems like a fast compression methods with competitive compression ratio. What additional storage costs are required when a method (such as IPR-L) is deployed across various devices?

Q4: Why do IPR-I and IPR-L appear to have similar results? Can these two approaches be unified into a single approach?

4 comments

Response to Reviewer s1di

Overall Response: We thank the reviewer for recognizing the decoding speed and cross-device evaluation. We clarify the scope of the speed claim, the conditions for applying IPR to another codec, and how the two forms share a common coding pipeline. All changes in the revised manuscript are highlighted in blue.


Q1/W1. Throughput and compression rate

TL;DR At comparable bitrates, IPR-L achieves the highest encoding and decoding throughput among the settings evaluated on Kodak on our GPU server.

Thank you for prompting us to clarify the scope of our throughput claim. Below are representative operating points from Appendix Table 9.

Codec / settingKodak rate (bpsp) ↓GPU-server encoding (MB/s) ↑GPU-server decoding (MB/s) ↑
JPEG XL, effort 23.4368147.3561.3
WebP, z63.203121.00184.0
IPR-L3.1937656.5826.4
JPEG XL, effort 33.146490.83237.1
JPEG XL, effort 72.90202.350164.4

IPR-L is faster than all evaluated learned codecs (Table 1); among classical codecs, the answer depends on the rate. We agree that some settings achieve lower bitrates, as JPEG XL effort 7 shows. The full sweep also includes JPEG XL effort 1, which encodes faster than IPR-L at a higher bitrate. At bitrates close to IPR-L’s, such as those of WebP z6 and JPEG XL effort 3, IPR-L is faster in both encoding and decoding. We added “at comparable rates” to the speedup claims in the abstract and Section 4.3 to make this scope explicit.


Q2. Transfer to existing codecs

TL;DR IPR is not yet a plug-in module, but its core idea is not tied to our pipeline, and we are interested in extending it to existing codecs in a backward-compatible way.

Thank you for this insightful question. We agree that compatibility (i.e. plug-in) with existing codecs is an important engineering consideration in practice.

IPR currently defines its own bitstream, so it is not a plug-in module for existing codecs. However, its core idea, carrying a compact probability model in the file and compiling it into a lookup table, is not tied to our pipeline. We are interested in extending IPR in a backward-compatible way. Existing codecs with binary context coding could then benefit from image-specific probability models. We see this as a natural next step and appreciate the suggestion.


Q3. Storage costs across devices

TL;DR Only IPR-L requires an installed asset: one ~7 MB LUT per device, shared across all files. IPR-I and IPR-L+MDL install nothing and carry their model or tree in each file.

The additional storage required on a device is as follows:

ModeAdditional asset
IPR-INone
IPR-L∼\sim7 MB LUT
IPR-L+MDLNone

The LUT is installed once and reused for every image, so it is not part of the file size.

Table 3 also reports parameters and runtime memory (excerpt; 512 × 512 images):

ModeInstalled assetsParametersGPU-server peak memory, Enc./Dec. (MB)macOS CPU peak RSS, Enc./Dec. (MB)
IPR-I011,153585.3 / 683.7 (VRAM)80.9 / 244.9
IPR-L7.0 MB010.9 / 10.6 (RSS)25.1 / 20.9
IPR-L+MDL0030.0 / 19.4 (RSS)49.8 / 49.5

The in-file model or tree is already counted in all reported rates. At runtime, the LUT-based forms stay within 50 MB of CPU memory.

IPR-L: context-width sweep on Kodak.

Context bitsRate (bpsp) ↓Dense LUT size (MB)
93.30810.44
113.21701.76
13 (default)3.19387.03
153.181928.11
173.1739112.46

The LUT size is a design choice rather than a fixed cost. The sweep above, taken from Appendix Fig. 9(a), varies the causal-context width (N(bi)\mathcal{N}(b_i)): each additional bit doubles the table, while the rate gain diminishes. The default of 13 bits keeps the table near 7 MB; a device with tighter storage can use an 11-bit table of 1.76 MB at a 0.7% higher rate.


Q4. Similar results of IPR-I and IPR-L, and unification

TL;DR IPR-I and IPR-L are two forms of one codec for different deployment paths (Fig. 7). Their rates are close because they share the coding pipeline; they differ in encoding cost and installed assets.

The new Fig. 7 shows how the forms work together in a single system: a server delivers IPR-I files, a device with an installed LUT codes locally with IPR-L, and devices share IPR-L+MDL files without requiring the LUT at the receiver.

Why the rates are similar. Both forms use the same reversible transforms, coding state, and range coder, and the IPR-L table is built from IPR-I probability models. Their rates therefore differ by only 1.7–5.3% (Fig. 5).

Where they differ.

Rate on Kodak (bpsp)Encoding (MB/s, GPU server)Installed assetPer-image fitting
IPR-I3.140.268NoneYes
IPR-L3.19656.57.0 MB LUTNo

Unification. The forms are already one codec: they share the transform, payload format, and decoding kernel, and differ only in where the probability table comes from.

Thanks for the response. My concerns have been addressed. I will keep my current score.

We are glad that our response has addressed your concerns. Thank you again for your constructive comments and review.

Reviewer_ihP4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes several statistical lossless image compression methods, with emphasis on fast encoding/decoding.

Critical strengths

This paper introduces a compact representation of image probability that is embedded in a bitstream. The method is low-complexity with good RD performance, compared to conventional compression methods.

Critical weaknesses
  1. The position of the proposed scheme is vague. IPR-L is a non-learning statistical method, achieving a compression rate comparable to JPEG 2000. Hence, comparing encoding/decoding to learning-based schemes can be pointless. IPR-I is an INR-based method and should be compared to its INR counterparts, instead of general methods. In conclusion, the numbers that this paper presents are misleading.

  2. Poor presentation of results. The paper is a little messy. Issues include, but are not limited to: Figure 5 has a clearly misleading scale (similar results but a significant visulized gap); multiple symbols/notations in Table 4 are not defined; the extension to lossy compression seems not very meaningful except serving as another ablation study.

  3. The formatting is not compliant with ICLR standards. Tables 1/4 and Figures 1/2/3 have incorrect small margins (there is a risk of desk rejection), where ICLR standards require leaving at least one line of space in between.

To conclude, the existing results of IPR-L show that it might be a competitive alternative among statistical methods, while the effectiveness of IPR-I is unclear without comparison with INR methods. The inappropriate experimental comparisons and poor organization substantially undermine the credibility of the reported results. I therefore recommend that the authors thoroughly restructure the manuscript, conduct fair and comprehensive evaluations against appropriate baselines.

Questions for authors
  1. Could the authors clarify the positioning of IPR-L and IPR-I and provide fair comparisons against appropriate baselines, particularly statistical compression methods for IPR-L and INR-based codecs for IPR-I?

  2. What is the actual technical significance of extending the proposed framework to lossy compression? Beyond serving as an additional ablation study, what new insights or advantages does this extension offer?

6 comments

Response to Reviewer ihP4

Overall Response: We thank the reviewer for identifying where the positioning and presentation need to be clearer. We clarify the deployment goals and comparison baselines for each IPR form, explain what the lossy extension demonstrates, and address the figure, notation, and layout concerns. All changes in the revised manuscript are highlighted in blue.


Q1/W1/W4. Positioning and baseline selection

TL;DR IPR-L is 4.0×/5.7× faster than classical codecs at comparable rates. For IPR-I, a controlled lossless INR comparison is not available, so we compare it with FNLIC and report adapted INR tests for reference.

Design rationale. IPR aims to make learned probability models deployable by removing two bottlenecks: (1) installing a shared pre-trained network, and (2) running neural inference repeatedly during coding. IPR-I removes both at the expense of per-image fitting; IPR-L and IPR-L + MDL remove this remaining bottleneck for on-device use:

  • IPR-I: Server-to-device delivery. A server fits the model once per image and embeds it in the file. The receiver compiles it into an image-specific LUT once and decodes by table lookup, with no installed model. It achieves the lowest rate among the three forms.

  • IPR-L: Repeated on-device coding. Per-image fitting is too costly for encoding on a device. IPR-L combines the IPR-I probability models into a reusable LUT and encodes and decodes images without fitting. The IPR-I prior improves the rate over count-only estimation on all four datasets (Appendix Table 10, Fig. 10).

  • IPR-L + MDL: Cross-device sharing. An IPR-L file can be decoded only with the same LUT. IPR-L + MDL instead carries a compact probability tree, so a receiver can decode it without the LUT.

To make these scenarios easier to follow, we added a figure to Section 5. Discussion, "Deployment scenarios" (Fig. 7).


IPR-L vs. statistical methods. As requested, we compare IPR-L with classical codecs across operating points (Appendix Table 9).

Codec / settingRate (bpsp) ↓Encoding (MB/s) ↑Decoding (MB/s) ↑
JPEG XL, effort 13.64181,109786.7
WebP, z63.203121.00184.0
IPR-L3.1937656.5826.4
JPEG XL, effort 33.146490.83237.1

No classical setting outperforms IPR-L in rate, encoding, and decoding throughput simultaneously. Accordingly, we revised the abstract and Section 4.3 to state the classical speedups (4.0×/5.7×) as relative to classical codecs at comparable rates.

IPR-I vs. INR-based codecs. The closest fitted lossless codec, FNLIC, is compared in Table 1 (IPR-I: 9% higher rate, ~1,600× faster decoding). IPR-I fits and transmits a model of conditional bit probabilities, whereas INR codecs (such as [A,B]) fit the signal itself. The official compression code of LosslessINR [A] is not available, so we list its reported rates on MNIST and pair its bit-plane INR with range coding on natural images. We evaluate the lossy INR codec PIC [B] under a file-size budget of 9.42 RGB bpp, set by IPR-I’s lossless rate on Kodak24.

Small images, for reference only (bpp). The original test indices of [A] are unavailable, so IPR-I is measured on a different set of 100 images (one model shared per file):

DatasetLosslessINR [A] (reported)IPR-I (measured)
MNIST2.581.86
Fashion-MNIST6.116.76

Natural images, lossless (bpsp). Our adaptation of [A]: a bit-plane INR fitted per image (SIREN, width 64) with its outputs range-coded; all bytes for independent decoding are counted:

DatasetLosslessINR [A] + range coding (our adaptation)IPR-I
Kodak246.113.14
TESTIMAGES40 (16-bit)13.899.71

Kodak24, PIC fitted per image within IPR-I's lossless file size (9.42 RGB bpp):

MethodRate (RGB bpp)PSNR (dB) ↑Lossless
PIC [B]1.5836.070/24
Enlarged PIC [B] (our adaptation)7.2636.930/24

In our tests, the bit-plane INR required 1.43–1.95× IPR-I’s rate for exact reconstruction, and neither PIC configuration reconstructed any Kodak image exactly within the 9.42-bpp budget.

[A] W. K. Han et al., "Towards Lossless Implicit Neural Representation via Bit Plane Decomposition," CVPR 2025.

[B] X. Liu et al., "PIC: Revisiting INR for Image Coding with Fast Encoding and Sub-Millisecond Decoding," ECCV 2026.

Q2/W2. Technical significance of the lossy extension

TL;DR The probability model already in the bitstream also improves lossy reconstruction, without extra transmitted bits or a separate reconstruction network.

Many applications need both exact originals for archiving and lower-rate versions for preview and transmission. Medical and space-borne imaging standards, such as DICOM [C] (which adopts JPEG 2000 [E]) and CCSDS 122.0 [D], support both modes for this reason. Dropping lower bit-planes lets IPR serve both with the same codec: lossy decoding reuses the same transform, probability model, and coder, and the rate is set by the number of omitted planes without retraining.

Beyond this, the IPR decoder reuses the transmitted probability model to estimate the omitted bits. On Kodak, this improves PSNR by up to 2.6 dB (Fig. 6), with the largest gain at low rates where truncation is coarsest. We view this not as a component ablation but as evidence that the probability model carried in the file captures signal structure beyond coding probabilities, so one model serves both entropy coding and reconstruction.

In Section 5 — Lossy extension, we added two sentences on the use of both modes and the reuse of the transmitted model, and revised the final sentence accordingly.

[C] Mildenberger, Peter, Marco Eichelberg, and Eric Martin. "Introduction to the DICOM standard." European radiology 12.4 (2002): 920-927.

[D] Yeh, Pen-Shu, et al. "The new CCSDS image compression recommendation." 2005 IEEE Aerospace Conference. IEEE, 2005.

[E] Marcellin, Michael W., et al. "An overview of JPEG-2000." Proceedings DCC 2000. Data compression conference. IEEE, 2000.


W2/W3/W4. Presentation and formatting

TL;DR We clarify how to read the bitstream figure, define the ablation notation, and address the flagged spacing.

  • Fig. 5 — Section 4.3, Representation cost and savings We redrew the figure as changes relative to IPR-L. The bars distinguish savings in coded bit-planes from added representation costs, and the diamonds show the net file-size change. Relative to IPR-L, IPR-I spends 0.4–1.3% of the rate on its representation and saves 2.3–6.1% in coded bit-planes, reducing the total rate by 1.7–5.3%; the savings exceed the cost by 2.4–7.4× across datasets. We now report these values explicitly instead of relying on the visual scale. The absolute bpsp composition from the original figure is moved to Appendix Fig. 11.

  • Table 4 — Section 4.3, Ablation study We define the symbols and identify what each ablation removes or replaces in the caption.

  • Tables 1,4 and Figs. 1–3 Fixed

  • Organization — Section 5, Deployment scenarios The new figure and accompanying text connect each coding form to its intended deployment setting.

Reviewer_ihP4OpenReview ↗

Discussion and response

I appreciate the effort the authors put to improve the paper. My minor concerns (W2 and W3) have been effectively resolved. One of my major concerns, that was, the position of the proposed scheme, is now clearer with the inclusion of Figure 7. My remaining major concern in W1(second part)/Q1 about the misleading numbers, however, is not resolved by the current response and revision.

Minor: Structure of the paper

The claim of the paper is too diverse and it makes the reader easy to lose the purpose of the paper. The message that current paper is conveyed is that a family of schemes are first proposed and evaluated, then finally assembled to form a system in Discussion and Figure 7. It is marginally acceptable for such a purpose, but the authors should first define the scenario that their schemes are designed for, then report fair comparison results. In the current paper, this has been done reversely, making readers hard to interpret the results before reading onwards.

Additionally, since authors now have one extra page in their budget, please avoid using \vspace or similar commands to override the submission template, particularly around the figures/tables here, which make the entire article look very dense. (I. think this issue is commented previously) 

Major: Inappropriate comparison with learning-based method and misleading wordings

The revised paper is still heavily misleading. For example, in Table 1/3, variants of IPR are compared to a number of learning-based methods. However, all proposed variants consistently under-performed the learning-based methods except L3C/FSAR in compression rate on DIV2K. For ArIB-BPS and OmniZip-L, IPR family even used 20% more bits. Hence, the comparison is not under equal or comparible performance, and is not imformative. While it is common knowledge that the main advantage and motivation of learning-based methods is great compression performance, degrading compression performance to conventional method level to regain encoding/decoding speed moves the IPR family away from existing learning-based methods. It is therefore deeply misleading to say that IPR makes learned probability models deployable, since it neither improves existing learning-based methods nor provides model-transferable designs.

From the results of Table 9, I recognise the value of IPR-L, since it encodes/decodes faster than conventional schemes, under comparable rate. However, the authors must amend the expression and stop claiming that they gain speedup compared to learning-based method, unless they can establish a fair comparison with comparable compression performance.

In conclusion, I am convinced that the proposed methods achieve speedup with the specified compression performance. However, the overclaims and misleading numbers forbid me from increasing the score for the paper at the current state.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.