- Rating
- 2
- Confidence
- 4
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper proposes HyPIC, a progressive image compression framework that combines Residual Vector Quantization (RVQ) with a novel Subspace-Projected Refinement (SPR) stage. The core idea is to follow N RVQ stages (which establish a shared joint latent prefix via learned codebooks) with a scalar-quantized correction in a learned low-dimensional subspace. The authors motivate this design with a local first-order analysis showing that residual refinement value depends on (i) alignment with the remaining decoder-space error and (ii) the decoded effect of the correction, neither of which is captured by Euclidean residual approximation alone.
The method is evaluated on Kodak, Tecnick, DIV2K, and CLIC 2020, reporting strong perceptual BD-rate gains over HiFiC, MS-ILLM, Control-GIC, PerCo, CTC, CMIC, RDVQ, and ProGIC, with competitive encoding/decoding speed (5.59 ms / 6.39 ms on 512×768 images). Ablations cover VQ depth, activation-free mappings, correction width, comparison against PCA/random/fixed-channel projections, and statistical treatments.
Critical strengths
-
This paper has a clear and well-motivated design analysis. Section 3.1 and Appendix B provide a clean local decomposition (Eq. 13) separating unreachable feature error from decoder-weighted coefficient error. This gives a principled rationale for learning synthesis directions (B) rather than simply coding the residual, and the distinction is empirically supported by Table 4: decoding the exact encoder latent y (PSNR 22.59, LPIPS 0.1910) is worse than the quantized SPR correction (PSNR 26.02, LPIPS 0.1440). This "exact latent recovery is not optimal" observation is a genuinely useful finding for the community.
-
This manuscript has strong empirical results across multiple datasets. The BD-rate reductions relative to Control-GIC are substantial (LPIPS: −72.08% Kodak, −64.08% Tecnick, −74.53% DIV2K, −75.46% CLIC). Figure 3 shows consistent dominance across all four datasets. The results are notable given the very low rate regime (0.03–0.20 bpp).
-
The paer shows excellent computational efficiency that is a meaningful practical contribution.
Critical weaknesses
1、While the paper compares against PCA and random projections (Figure 6), there is no comparison against a well-tuned scalar residual coder (e.g., a hyperprior-based SQ codec operating directly on the 256-dim residual). The "5RVQ bypass" in Table 2 uses the unchanged VQ decoder, which is not a fair SQ baseline. I want to see whether SPR's gains persist against a properly trained SQ refinement stage. The claim that "fixed-dimensional reductions do not reproduce the benefit of learned directions" is supported, but the comparison set could be stronger.
2、The authors state: "As an official ProGIC implementation was unavailable for our evaluation, we use an in-house reimplementation to measure its computational cost on the same hardware." However, ProGIC's RD performance in Figure 3 appears to come from the original paper, while its computational cost comes from the reimplementation. This mixed protocol should be clarified — are the RD curves for ProGIC from the original paper or the reimplementation? If from the original paper, hardware/implementation differences could confound the speed comparison. If from the reimplementation, the RD curve may not reflect ProGIC's true performance.
3、Table 7 uses a different configuration than the main model. The width study uses four-VQ variants trained for only 100,000 updates, while the main model uses three VQ stages trained for 1.5M updates. The statement "We use d = 10 in the three-VQ main model as a practical design choice separately from this four-VQ width comparison" is honest, but the reader cannot tell whether d = 10 is actually near-optimal for the three-VQ model. A width sweep on the same configuration as the main results would be more informative.
4、The paper claims "four progressively decodable operating points from one model," but there is no figure or table showing the rate-distortion trajectory as stages are added (e.g., PSNR/LPIPS at each prefix vs. cumulative bpp). Figure 5 shows the effect of VQ depth, but it appears to compare separate models rather than the progressive decoding of a single model. A plot showing the quality at each operating point (prefix 1, prefix 2, prefix 3, final SPR) would substantiate the progressive claim.
5、Entropy model details are underspecified. Eq. (6) combines a Gaussian score and learned discrete logits, but the paper does not specify:
6、How the Gaussian score is computed (is it a learned Gaussian over codewords? A distance-based score?).
7、Whether the discrete logits are spatial-context-dependent (the text mentions "top and left spatial context" but the architecture in Figure 2b is not fully clear).
8、How the hyperprior features f_k are combined with the codeword scores.
9、Equation numbering: Eq. (10) is referenced in the text but I could not find it in the provided pages (it may be in Appendix B; if so, the reference in Section 3.4 should point there).
10、Table 1 formatting: The table is difficult to parse in the provided PDF (columns run together). Ensure the final version is legible.
11、Figure 1 caption: "Gold outlines mark selected cells" — but the figure shows red stars and dots; the description could be clearer.
12、Typos: "igure 1" (p. 2), "igure 3" (p. 7), "igure 5" (p. 9) — missing "F". "Balle et al., 2016" is misspelled as "Balle" (correct is Ballé). "Toderici et al., 2015" is cited but the reference list shows "Toderici et al., 2015" — consistent, but check the 2017 entry.
13、Reference [Cao et al., 2026]: ProGIC is cited as CVPR 2026; ensure this is a real, forthcoming publication (the year is in the future relative to typical review cycles — this may be a preprint or a placeholder).
14、ProGIC comparison: Are the RD curves in Figure 3 for ProGIC from the original paper or your reimplementation? If from the original paper, how do you ensure the speed comparison in Table 1 is fair?
15、SQ baseline: How does HyPIC compare against a hyperprior-based SQ codec operating on the full 256-dim residual at the same total rate? This would isolate the benefit of the learned subspace from the benefit of scalar quantization itself.
16、Progressive decoding: Can you provide a plot of PSNR/LPIPS vs. cumulative bpp for each of the four operating points from a single model, to substantiate the progressive claim?
17、Activation-free mappings: Why do identity maps outperform SiLU/ReLU? Is this specific to the SPR stage, or does it also hold for the VQ stages?
18、Width selection: Is d = 10 optimal for the three-VQ main model? What happens at d = 4, 8, 16 for the same configuration?
19、Entropy model: Can you provide the exact architecture of the Gaussian score and discrete logit branches in Eq. (6), and the parameter head for δ and σ?
20、Generalization: How does HyPIC perform on out-of-distribution content (e.g., medical images, screenshots, text)? The training set is ImageNet-1k, which may bias toward natural images.
21、Rate range: The evaluated rate range is 0.03–0.20 bpp. How does HyPIC perform at higher rates (e.g., 0.5–1.0 bpp)? Does the advantage persist, or does it diminish?
Questions for authors
Please see the weaknesses part
Dear Reviewer Li4J,
Thank you again for your detailed review and for recognizing the motivation behind learning correction directions. We submitted our rebuttal about five days ago, grouping your 21 questions into related topics while retaining their original numbering. With three days remaining in the discussion period, we would sincerely appreciate knowing whether our responses have resolved your concerns.
The main points of our response are:
We value the constructive discussion that is an important part of ICLR’s review process and sincerely hope to continue it with you. Please let us know which points, if any, would benefit from further explanation; we are ready to respond promptly during the remaining discussion period.
Thank you for your time and consideration.
Best regards,
The Authors