- Rating
- 3
- Confidence
- 4
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper addresses how to improve restoration fidelity with ultra-low bit widths in post-training quantization (PTQ). Through controlled analysis, it observed a key characteristic that some special outliers, named as pedestal, define a direction highly sensitive to quantization error, making it a meaningful structure. Practical strategies, including Pedestal Shifting for pedestal isolation and Grill for quantization bounds calibration are proposed. Experiments are extensive, showing consistent and nontrivial improvement over existing methods.
Critical strengths
-
Regarding pedestal, the paper provides compelling empirical evidence that a persistent dominant activation component is approximately constant across tokens yet highly sensitive to quantization-induced distortion.
-
The PSG strategies, including Pedestal Shifting and Grill, are well grounded and effective.
-
Experiments, both quantitative and qualitative, are extensive and demonstrate substantial improvements, particularly at ultra-low bit widths.
Critical weaknesses
-
The mechanistic evidence described in Sec.3 and Figure 1 is uneven across architectures and datasets. Although Sec. 3.2 reports directional perturbation results averaged across Mamba-based restorers, the detailed token-wise constancy experiments in Appendix B are restricted to MambaIR ×2 on Set5, and the actual quantization-error decomposition is reported only for MambaIR. Extending these analyses to other backbones and datasets would strengthen the generality of the proposed explanation.
-
The baseline results exhibit an unexplained non-monotonic trend across bit widths. In Table 8, SPR²Q at W3A3 outperforms its W4A4 counterpart on all five benchmarks for MambaIR ×2, including gains of 0.557 dB on Urban100 and 0.715 dB on Manga109. This does not necessarily indicate an error, but warrants clarification of the baseline reproduction protocol.
Questions for authors
-
Could the authors extend the token-wise constancy and quantization-error decomposition analyses to additional backbones and datasets?
-
Could the authors explain the non-monotonic SPR²Q results in Table 8?
Given satisfactory clarifications of the two points above, the reviewer is willing to raise the score further.
3 comments
Response to Reviewer WKAV (2/2): Non-monotonic SPR²Q results in Table 8 (W2, Q2)
Summary. We verified our SPR²Q setup by reproducing the reported ×2 results on MambaIRv2-light under the original protocol (Table R5). On MambaIR ×2, the W4A4 model degrades during distillation under the released learning rate. Reducing only the learning rates by 10× restores W4A4 above W3A3 on all five benchmarks, while PSG remains ahead by at least 0.66 dB. The same reduction does not improve W3A3 or W2A2.
Protocol: official code, released settings.
For MambaIR, we retain the optimization settings released for SPR²Q's MambaIRv2-light backbone: Adam, learning rate 10⁻², batch 8, 12,000 + 500 iterations, at every bit width. The released configuration keeps the rate constant, whereas the paper states cosine annealing. We follow the release, as for every baseline. Only two things differ from it, both shared by all methods: the quantization scope (Sec. 5.1) and the use of the final checkpoint (response to Reviewer 2JMm). With its original scope and checkpoint selection, the same code reproduces the reported MambaIRv2-light results (Table R5).
SPR²Q's reported evaluations and released configurations cover lightweight backbones. On MambaIRv2-light, SPR²Q is the strongest baseline at W4A4 at every scale of our tables, with W4A4 above W3A3 (Tables 2 and 9). MambaIR is substantially larger (82.0 MB versus 3.4 MB, Table 4(a)), and no configuration is released for it. We therefore transferred the released settings without backbone-specific tuning.
W4A4 degradation on MambaIR ×2 is resolved by reducing the learning rate.
The official code logs Set5 PSNR during distillation. The W4A4 model drops 0.49 dB below its calibrated start within the first 1,000 iterations, never fully recovers, and ends 0.19 dB below it.
Under the same rate, more weights move to a different quantization level at W4A4 (Table R10). SPR²Q trains low-rank rectifiers (1.66M parameters on MambaIR) jointly with the quantizer bounds. Their weight increments are added to the weights before quantization. The trained increments have a similar size relative to the weights at both bit widths. On the finer 4-bit grid, they move 24.2% of the weights to a different level, against 10.9% at W3A3. During distillation, Set5 PSNR rises by 1.16 dB at W3A3 and falls by 0.19 dB at W4A4. These observations led us to test whether a rate that suits W3A3 is too large at W4A4.
| SPR²Q on MambaIR ×2, released rate | W3A3 | W4A4 |
|---|---|---|
| Rectifier weight increment, relative to weight norm | 8.5% | 8.2% |
| Weights moved to a different quantization level | 10.9% | 24.2% |
| Change in Set5 PSNR during distillation (dB) | +1.16 | −0.19 |
Table R10. Statistics of the final checkpoints.
Reducing the learning rates by 10× restores the expected ordering (Table R11). We re-ran W4A4 with the learning rates divided by 10 and nothing else changed. The final checkpoint now improves on its calibrated start, and W4A4 is above W3A3 on all five benchmarks. PSG leads this row by at least 0.66 dB and every other baseline of this setting by at least 0.32 dB, so the conclusion of Table 8 does not depend on it.
| MambaIR ×2, PSNR (dB) | Set5 | Set14 | B100 | Urban100 | Manga109 |
|---|---|---|---|---|---|
| SPR²Q W3A3, released rate | 36.961 | 32.659 | 31.444 | 30.015 | 36.657 |
| SPR²Q W4A4, released rate | 36.794 | 32.515 | 31.317 | 29.458 | 35.942 |
| SPR²Q W4A4, rate ÷ 10 | 37.210 | 33.083 | 31.709 | 31.042 | 37.507 |
| PSG W4A4 | 38.202 | 34.104 | 32.370 | 33.225 | 39.191 |
Table R11. Bold: the re-run with the smaller rate.
The same reduction does not improve W3A3 or W2A2 (Table R12). The smaller rate lowers W3A3 by up to 1.39 dB and W2A2 by less than 0.05 dB. We therefore retain the originally reported W3A3 and W2A2 results.
| ×2, change with rate ÷ 10 (dB) | Set5 | Set14 | B100 | Urban100 | Manga109 |
|---|---|---|---|---|---|
| W4A4 | +0.416 | +0.568 | +0.392 | +1.584 | +1.565 |
| W3A3 | −0.814 | −0.432 | −0.361 | −0.887 | −1.393 |
| W2A2 | −0.045 | −0.018 | −0.013 | −0.009 | −0.026 |
Table R12. SPR²Q on MambaIR ×2.
PSG uses one setting everywhere.
It trains only the bounds, with one setting for every backbone, task, and bit width (App. C).
Revision (done). Table 12 (Table 8 of the submission) now reports the reduced-rate SPR²Q W4A4 row with SSIM, marked †, and Appendix C states the setting and the reason. We also ran cosine annealing: the W4A4 model degraded again, so it is not reported.
Revised PDF uploaded.
We have uploaded the revised paper (changes in blue) and updated the "Revision" paragraphs of our two responses above with the locations. The extended analyses are in Appendix B.1 and B.2 (Tables 8 and 9), and the updated SPR²Q row is in Table 12 (Table 8 of the submission). We hope the revision addresses your concerns and would appreciate it if you could take a look.
Response to Reviewer WKAV (1/2): Constancy and error decomposition beyond MambaIR (W1, Q1)
We thank the reviewer for describing the pedestal as "a meaningful structure" supported by "compelling empirical evidence", and Pedestal Shifting and GRILL as "well grounded and effective". This is the chain the paper is built on: the analysis identifies a structure that should be preserved, and the method is designed around preserving it. As requested, we extended the two analyses behind this structure, token-wise constancy and quantization-error decomposition, to more backbones, tasks, and test sets.
Summary. Together, the two analyses now cover four backbones, including the Transformer, three tasks, and seven test sets. The findings of Sec. 3 carry over. The response along e1 stays functionally constant at the gated sites, and the e1-aligned error stays a small part of the distortion with a disproportionate effect on fidelity.
Token-wise constancy (Table R8).
We repeated the joint replacement of Appendix B on the full-precision models, where the token-dependent coefficient ct along e1 is replaced by a single constant (Eq. 2). The gate uses the same threshold as in the paper (τ=0.1).
Table R8. PSNR change (dB) under joint replacement of ct by a single constant. The first row is the setting of Table 7(a). MambaIRv2-light is released for SR only.
Quantization-error decomposition (Table R9).
We repeated the decomposition of Sec. 3.2 on the three other backbones. The activation quantization error is split into its component along e1 (ϵ∥) and the rest (ϵ⊥), as in Eq. 3, and each is removed in turn.
Table R9. Recovered: share of the lost fidelity restored when one error component is removed. MambaIR is reported in Sec. 3.2.
Revision (done). Both analyses are in Appendix B.1–B.2; Sec. 3.1 and 3.2 refer to them.