- 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
- 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.
- 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
-
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.
-
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 appears inconsistent with Equation (4).
- In Line 228, the citation of KT smoothing is not separated from the main text.
Questions for authors
-
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?
-
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
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.
Response to Reviewer 5p7y
Q1/W1. Contribution relative to prior work
Thank you for raising this point. As the reviewer notes, IPR builds on established coding techniques. Our design starts from a practical question:
This imposes two requirements:
The decoder should not need a shared pre-trained model.
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.
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
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.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%.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
Both baselines use the same state s, 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), but differ in how they obtain the counts:
For the reusable IPR-L table,
Appendix Table 10andFig. 10compare 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.