- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes WideR, a neural image compression framework that supports a wide bitrate range and distortion-perception trade-off control. Through rate-adaptive loss weights, the authors emphasize perceptual quality at low bitrates and distortion optimization at high bitrates. Combined with a one-step diffusion prior and decoder-side LoRA interpolation, this enables reconstruction preference control using the same bitstream.
Critical strengths
- A single model supports bitrate control over a wide range of approximately 0.02-1.0 bpp, along with decoder-side distortion-perception adjustment, which has some practical value.
- The experiments are relatively extensive, including endpoint performance, comparisons at matched PSNR/FID, and ablation studies.
Critical weaknesses
- The core weight schedule remains relatively empirical, lacking sensitivity analyses of the boundary weights and scheduling forms. The overall approach is very simple.
- The fixed-weight-ratio ablation only uses a ratio of 4 without separate tuning. The validation of the advantages of adaptive scheduling is too limited. This is the core method.
- The conclusion about representation limitations relies on an experimental setting where the pretrained backbone is frozen and only modules such as LoRA are trained. This is insufficient, and I need to see more justification, especially since full fine-tuning is not performed.
- The model is large, with 864M total parameters, significantly more than EGIC's 33M. The decoding latency with the diffusion branch enabled is 179 ms, also higher than EGIC's 108 ms. This cost is too high.
- More comparisons should be provided on LPIPS and DISTS.
Questions for authors
- Do the advantages of adaptive scheduling remain stable when changing the boundary weights, adopting other scheduling forms, or thoroughly tuning the fixed weight ratio? I think these experiments can be added quickly.
- Do the observed cross-endpoint performance gaps still exist after increasing the LoRA rank or fully fine-tuning the decoder?
- Could you provide peak GPU memory usage and decoding latency for high-resolution images to help us evaluate the actual deployment overhead?
- Is the FID BD-rate as high as 60%? Is there any performance loss in certain cases?
- How about LPIPS and DISTS RD performance?
5 comments
Response to Reviewer PVAb [2/2]
Response to Questions
1. Stability of the Advantages of Rate-Adaptive Balancing
[Discussed in Weaknesses 1–2] As long as the monotonic increase of the weight ratio is kept, the advantage of WideR is not sensitive to the boundary weights or the interpolation function (Tables 3–4). In addition, the fixed ratio of 4, the midpoint of our rate-adaptive ratio, is clearly outperformed by rate-adaptive balancing (Section 5.4).
2. Fully Fine-Tuning the Decoder
The backbone is frozen identically in all experiments of DoR, PoR, and WideR to preserve the prior knowledge of the perceptually pretrained VAE decoder. This is a controlled experiment, in which the three models differ only in their objectives. To further examine the representation capacity in a more extreme setting, as you suggested, we add an experiment that fully fine-tunes the VAE decoder during the training of DoR and PoR (the new Table 13, Appendix M). On Kodak24, full fine-tuning does not reduce the perception penalty of DoR (+2.3%) and reduces the distortion penalty of PoR by only 14.8%, whereas rate-adaptive balancing (ours) reduces them by 68.0% and 90.3% with the decoder frozen. Therefore, the cross-endpoint penalties do not come from the frozen backbone but from the training objective.
| Kodak24 | Perception penalty (ΔFID) ↓ | Distortion penalty (ΔPSNR, dB) ↓ |
|---|---|---|
| DoR | 9.64 | – |
| DoR (w/ VAE decoder full fine-tuning) | 9.86 (+2.3%) | – |
| PoR | – | 1.74 |
| PoR (w/ VAE decoder full fine-tuning) | – | 1.48 (−14.8%) |
| Rate-adaptive balancing | 3.08 (−68.0%) | 0.17 (−90.3%) |
3. Peak GPU Memory Usage and Decoding Latency for High-Resolution Images
We agree that deployment feasibility at high resolutions is important. We add the peak GPU memory and encoding and decoding latency at the DIV2K resolution (2040×1356) to the new Appendix N (Table 14) and report them below.
| 2040×1356 (DIV2K) | Latency, enc. / dec. ↓ | Peak memory, enc. / dec. ↓ |
|---|---|---|
| EGIC | 680 / 853 ms | 4.2 / 8.2 GB |
| AEIC-ME | 122 / 4,333 ms | 5.7 / 9.5 GB |
| OneDC | 1,573 / 6,998 ms | 13.2 / 16.9 GB |
| WideR | 322 / 4,641 ms | 5.9 / 9.6 GB |
(All models decode the full image in a single pass without tiling.)
At this resolution, WideR decodes slower than EGIC (4.6 vs. 0.9 s) and requires more peak GPU memory in both the encoder and the decoder. The slower decoding mainly comes from the attention modules of the pretrained diffusion model, whose cost grows quadratically with the image size, and AEIC-ME and OneDC show a similar trend.
Please note that EGIC does not consider ultra-low rates and thus does not use a large generative prior. In contrast, WideR also targets ultra-low rates, where most details are smoothed out or lost in compression, and a large diffusion prior is effective in naturally generating such missing information. Generative codecs dedicated to ultra-low rates also use such a prior, and their complexity is similar to that of WideR (Table 5). At 2040×1356, AEIC-ME and OneDC also take several seconds to decode and need similar or more decoding memory than WideR, as shown above. Thus, we believe that the complexity of WideR is better compared with that of EGIC and an ultra-low-rate generative codec together, rather than with EGIC alone.
In addition, our rate-adaptive balancing itself does not depend on a specific diffusion prior. Thus, we believe that applying it to lightweight diffusion priors such as CoD-Lite (Jia et al., 2026), which are actively studied, is a promising next step. As this study does not cover such an extension, we leave it as a limitation in Section 6.
4. FID BD-Rate of 60% and Performance Loss in Certain Cases
Thank you for the question. Yes, −60.0% is the P-endpoint FID BD-rate of WideR against EGIC on CLIC 2020 (Table 1; −50.3% on Kodak24 and −51.7% on DIV2K). This gain is consistent with the curves and the visual examples in Figures 1, 16, and 17, which show no performance loss against EGIC. WideR also achieves a better D-endpoint PSNR BD-rate than EGIC on all three datasets (Table 1).
5. LPIPS and DISTS RD Performance
Thank you for the suggestion. Table 1 reports the P-endpoint BD-rates in LPIPS and DISTS on all three datasets, and Figure 16 (Appendix I) shows the rate–LPIPS and rate–DISTS curves of all compared codecs. However, our ablations and matched-quality comparisons (Table 2) use only PSNR and FID. We would be glad to add these results in LPIPS and DISTS if you find them helpful.
We hope these responses address your concerns. If any point remains unclear, we would be glad to provide further clarification or results during the discussion period.
Thank you for the additional experiments. The sensitivity analyses and full fine-tuning result address some of my concerns. However, the comparison still uses only a fixed weight ratio of 4, with additional results pending. And different fixed ratios may perform differently across bitrates, so a more thorough comparison is necessary to demonstrate the benefit of adaptive weighting. The high-resolution decoding cost also remains substantial. So I will keep my original rating for now.
Response to Reviewer PVAb [1/2]
We greatly appreciate your insightful and detailed feedback. Your comments have been carefully discussed and fully taken into account, and we improve our work accordingly.
Addressing Weaknesses
Thank you for these important points. We add sensitivity analyses of the boundary weights and the interpolation function (we now use this term instead of "schedule" to avoid confusion) in the new Tables 3–4. The core of our method is the monotonic increase of the distortion-to-perception weight ratio with rate (Eq. 3), and the geometric interpolation of Eq. 4 is one natural instance. To show that its advantage is not sensitive to the choice of the other hyperparameters, both analyses keep this property and vary only the other choices. For the boundary weights, we fix λPlo and λDhi, which make the perception term dominant at the lowest rate and the distortion term at the highest rate and let a single model span a wide rate range. We then vary λDlo or λPhi, the weights of the opposite loss at each end, over several values (Table 3).
Changing the boundary weights alters the overall D–P balance and shifts the operating points, i.e., one endpoint improves while the other degrades. However, these shifts are small compared with the advantage of WideR over EGIC. Thus, as long as the monotonic increase is kept, the advantage of WideR is not sensitive to the boundary weights, and every variant achieves a D-endpoint performance similar to EGIC and a much better P-endpoint performance.
To see how the results change with the shape of the interpolation function, we replace q in Eq. 4 with qβ, which makes the log of the weight ratio linear (β = 1), convex (β = 2), or concave (β = 0.5) in q (Table 4).
Changing β also shifts the operating points and thus changes the BD-rates against the paper's setting. However, the BD-rates against EGIC show a different trend, because β shifts the operating points mainly at the lowest rates of WideR, which EGIC does not cover. Over the full rate range, β = 1 shows the most balanced performance between the two endpoints (Figure 11b in Appendix D). Every variant achieves a D-endpoint performance similar to EGIC and a much better P-endpoint performance.
These two analyses show that as long as the monotonic increase is kept, the advantage of WideR is not sensitive to the boundary weights or the interpolation function. We believe the simplicity of our approach is a strength, as it adds no parameters or inference cost yet substantially reduces the cross-endpoint penalties (Figure 4).
The purpose of this experiment is to show that jointly optimizing the D and P objectives alone is not sufficient, and that changing their balance with rate is necessary. The ratio of 4 is not a tuned value but the midpoint (q = 0.5) of our rate-adaptive ratio, i.e., the geometric mean of λD(q)/λP(q) under the uniform sampling of q (Appendix C). Thus, the fixed ratio and the rate-adaptive ratio have the same average balance and differ only in whether the ratio changes with rate. Even in this matched comparison, rate-adaptive balancing (ours) achieves a −18.1% PSNR BD-rate at matched FID and a −84.3% FID BD-rate at matched PSNR (Section 5.4). Nevertheless, we are training two more fixed ratios, 2−1.5 and 25.5, which equal the rate-adaptive ratio at q = 0.25 and 0.75. Their training is not expected to end within the rebuttal period, so we will add the results to Section 5.4 and Appendix C in the final version. Given the large gaps of the fixed ratio of 4, we strongly expect the same conclusion.
We answer this point in Question 2.
We answer this point in Question 3.
We answer this point in Question 5.