OSCAR: One-Step Diffusion Codec Across Multiple Bit-rates

Thanks for the thorough review and valuable comments. We are encouraged by your recognition of the elegance of our method and the clarity of our empirical results. Below we respond to your specific comments.

• Q1: The paragraph spacing issue.

  Thank you very much for your attention to this matter. We strictly followed the official LaTeX template, but acknowledge that some section and caption placements may have affected readability. We will carefully revise the formatting in our revised paper to ensure a clear presentation.

• Q2: Authors should expand the related work.

  Thanks for the suggestion. We will add the citations and include the comparison results in the revised paper, as well as a discussion of similar works in the related work section.

  • Comparison with DDCM compression approach.

    In our method, we use LPIPS-AlexNet as a training loss, and LPIPS-VGG as the evaluation metric to ensure a fair comparison. Here we follow DDCM to use LPIPS-AlexNet (denoted as LPIPS*) and crop DIV2K-val images to 768$\times$768 for evaluation.

    model	bpp	PSNR↑	LPIPS*↓	FID↓
    DDCM (s=1000)	0.0420	21.1670	0.2010	18.3300
    OSCAR (s=1)	0.0313	21.3304	0.1958	31.8673
    DDCM (s=1000)	0.0660	21.7800	0.1720	16.3960
    OSCAR (s=1)	0.0507	21.9478	0.1554	28.9917
    DDCM (s=1000)	0.1370	22.7070	0.1380	14.6620
    OSCAR (s=1)	0.1250	23.4529	0.1097	15.7407

    Compared to the 1000-step DDCM, our one-step OSCAR achieves better performance on reference-based metrics (PSNR and LPIPS*), indicating more accurate reconstructions. Although DDCM yields better FID scores, the performance gap narrows as the bitrate increases.

  • Comparison with Cold Diffusion.

    We carefully reviewed this paper and found that it focuses on speech enhancement, without an image compression module and an open-source implementation. We'd appreciate it if you could clarify how it might be adapted to image compression.

  • What is the PerCo Version? Did the authors finetune 0.051 bpp and 0.094 bpp PerCo themselves?

    We use PerCo (SD) as our baseline. For 0.0313 bpp and 0.1250 bpp, we use the checkpoints released in their open-source code. For 0.051 bpp and 0.094 bpp, we train the model ourselves using their codebase. We will clarify this in the revised manuscript and include the corresponding paper in our citations.

• Q3: Authors should supplement PSNR, MS-SSIM, and FID results, and discuss how the latent-space nature of the method hurts pixel-space distortion.

  Thank you for the valuable advice. We omitted the MS-SSIM results of BPG, DPICT, and CTC in the original paper as they generally underperform MS-ILLM, which we reported as a representative non-diffusion baseline. We will include the full results and the detailed discussion in our revised paper.

  Tables below present results on the DIV2K-val dataset.

  model	bpp	PSNR↑	MS‑SSIM↑	FID↓
  BPG	0.0584	23.9356	0.8624	127.6502
  DPICT	0.0745	24.5078	0.8896	78.0774
  CTC	0.0707	25.8485	0.9166	75.4264
  MS‑ILLM	0.0493	24.5745	0.8814	42.6882
  PerCo (s=20)	0.0507	19.0644	0.6883	29.8819
  DiffEIC (s=50)	0.0414	21.7711	0.7586	29.4854
  OSCAR (s=1)	0.0507	21.9662	0.7623	28.8701

  model	bpp	PSNR↑	MS‑SSIM↑	FID↓
  BPG	0.1361	26.1172	0.9293	94.8704
  DPICT	0.1401	26.3581	0.9336	41.4491
  CTC	0.1399	26.8273	0.9365	35.2881
  MS‑ILLM	0.1622	27.0104	0.9417	18.3835
  PerCo (s=20)	0.1250	21.9182	0.8613	16.8871
  DiffEIC (s=50)	0.1264	23.4237	0.9088	15.1083
  OSCAR (s=1)	0.1250	23.4810	0.8906	15.7285

  We observe that diffusion codecs generally outperform non-diffusion methods in perceptual metrics such as FID (shown in the table), LPIPS, and DISTS (reported in our main paper), but underperform in pixel-level fidelity metrics like PSNR and SSIM. OSCAR, along with PerCo, DiffEIC, and the recent DDCM, all exhibit this behavior. We attribute this common behavior to the inherent characteristics of the diffusion backbone.

  This trade-off between pixel-wise fidelity and perceptual quality arises mainly from two factors. (1) The strong generative prior in pre-trained diffusion models, which prioritizes perceptual quality over pixel-wise reconstruction. (2) The nature of denoising in latent space amplifies this trade-off, as the decoder must reconstruct images from compressed, high-level features, which are optimized for semantic or perceptual quality rather than pixel accuracy. Unlike pixel-space compression, latent-space representations inherently discard low-level details, making exact pixel-wise reconstruction more difficult.

  Notably, this gap in pixel-level fidelity tends to narrow at higher bitrates, especially for MS-SSIM. This suggests that the limitations of latent-space reconstruction can be gradually mitigated when more information is retained.

• Q4: The method supports multiple bitrates with one model, but is limited to pre-chosen bitrates.

  The goal of our method is not to support a very fine-grained range of bitrates, as in traditional variable-bitrate approaches. Rather, we aim to explore whether the proposed bitrate-to-timestep mapping can enable a one-step diffusion model to support a set of bitrates while achieving performance comparable to single-bitrate models. Our experiments provide strong evidence that OSCAR is effective in achieving this target.

  To the best of our knowledge, current variable-bitrate methods still underperform single-bitrate baselines (Figure 5 in our paper). While our approach is currently validated on a set of predefined bitrates, we believe it offers a promising path toward achieving both flexibility and high performance in future designs.

• Q5: The entire model needs to be retrained for every new wanted bitrates, while PerCo just finetunes a pre-trained encoder.

  We acknowledge that the diffusion backbone needs to be retrained for new bitrates. However, we demonstrate that OSCAR can generalize well to new bitrates with minimal computational overhead.

  We adopted a simple training scheduler: half of the training steps are allocated to the new bitrate, and the other half to the original bitrate group. Extending to a new bitrate requires only 15K training iterations, which take less than 4 hours on 4×A6000 GPUs. We attribute this efficiency to the fact that the diffusion model has already learned to handle structurally similar latent distributions, enabling fast and stable adaptation to nearby bitrates, while causing minimal performance degradation on previously trained bitrates.

  The tables below report results on the DIV2K-val dataset, where we progressively adapt the model to new bitrates one at a time. Numbers in parentheses indicate the performance change compared to the model before adaptation to the new bitrate. These results will be included in the revised paper.

  bpp	DISTS↓	LPIPS↓	MS-SSIM↑
  0.0313	0.1262	0.3883	0.7210
  0.0507	0.1102	0.3487	0.7623
  0.0937	0.0694	0.2385	0.8713
  0.1250	0.0617	0.2157	0.8906

  bpp	DISTS↓	LPIPS↓	MS-SSIM↑
  0.0313	0.1306 (↑3.49%)	0.3927 (↑1.13%)	0.7161 (↓0.68%)
  0.0507	0.1141 (↑3.54%)	0.3523 (↑1.04%)	0.7578 (↓0.59%)
  0.0937	0.0728 (↑4.90%)	0.2427 (↑1.77%)	0.8595 (↓1.35%)
  0.1250	0.0648 (↑4.93%)	0.2198 (↑1.91%)	0.8766 (↓1.57%)
  0.0430 (new)	0.1198	0.3670	0.7402

  bpp	DISTS↓	LPIPS↓	MS-SSIM↑
  0.0313	0.1360 (↑4.19%)	0.4006 (↑2.03%)	0.7031 (↓1.81%)
  0.0507	0.1185 (↑3.87%)	0.3589 (↑1.87%)	0.7459 (↓1.57%)
  0.0937	0.0752 (↑3.34%)	0.2457 (↑1.25%)	0.8506 (↓1.03%)
  0.1250	0.0672 (↑3.75%)	0.2228 (↑1.38%)	0.8664 (↓1.16%)
  0.0430	0.1246 (↑4.01%)	0.3742 (↑1.98%)	0.7282 (↓1.62%)
  0.0781 (new)	0.0918	0.2766	0.8412

  In addition, we carefully examined the training procedure of PerCo. In their original paper, they state that all linear layers in the UNet (about 15% of the model) must be finetuned for each new bitrate. In contrast, our method uses LoRA to finetune only 1.8% of the total parameters. PerCo (SD) finetunes the entire UNet and hyper-encoder according to the open-source implementation. We kindly request you to double-check this point.

• Q6: The visualization should include full images or large crop comparisons.

  Thank you for the suggestion. We have prepared the full image comparisons. However, we noticed that additional PDFs are not permitted in rebuttal this year. We will include these visual results in the revised version of our paper.

• Q7: How many parameters are added by the rank 16 LoRA?

  The SD-2.1 UNet has 882.1M parameters, and the rank-16 LoRA introduces an additional 16.2M parameters.

We hope this response can help address your concerns. We believe our work can have an important impact on the field of image compression by offering a new perspective on leveraging pre-trained diffusion models for efficient one-step reconstruction across multiple bitrates. If you find our response satisfactory, we would be sincerely grateful if you could consider a higher score, as it would greatly support our efforts. Thank you once again for your consideration.

Thank you very much for the response. We appreciate your inclination to increase the rating. We commit to further refining the spacing in line with the template to enhance readability, as well as appending the discussed matters to our revised paper. Below, we provide our detailed response to the remaining concern.

• Q1: Comparison with Cold Diffusion.

  Thank you for the clarification. Cold Diffusion is specifically designed for image restoration tasks such as deblurring, denoising, and super-resolution. However, based on both the paper and its open-source implementation, we do not find a straightforward way to adapt it for image compression. Nevertheless, the pixelated degradation idea is interesting and inspiring, and we will include this paper in the related work section of our paper.

• Q2: The reconstruction bound for VAE.

  We conducted an additional experiment by directly passing the original image through the SD-2.1 VAE. As shown in the table below, the pixel-wise fidelity of the VAE is relatively low—only comparable to CTC at 0.1399 bpp. This provides a clear explanation for why diffusion models tend to lag behind in pixel-level fidelity. We value this insightful suggestion and will include these results in the revised paper.

Thanks for the valuable comments, constructive advice, and recognition of our idea and empirical results. Below we respond to your specific comments.

• Q1: Visual comparisons with DiffEIC.

  Thank you for the suggestion. We have prepared the visual comparisons with DiffEIC. However, as additional PDFs are not permitted in rebuttal this year, we will include these visual results in the revised version of our paper.

• Q2: Does the proposed method show better fidelity than the existing diffusion-based method?

  Pre-trained diffusion models inherently prioritize perceptual quality over exact pixel-wise reconstruction due to their strong generative priors. As a result, diffusion codecs often produce visually appealing images at the cost of pixel-level fidelity. Since OSCAR also builds upon a diffusion backbone, this trade-off naturally applies. However, despite supporting multiple bitrates and requiring only a single denoising step, OSCAR achieves performance competitive with DiffEIC and significantly surpasses PerCo in both perceptual and pixel-level fidelity.

  The table below reports results on the DIV2K-val dataset:

  models	bpp	DISTS↓	LPIPS↓	MS-SSIM↑
  PerCo (s=20)	0.0507	0.1647	0.3646	0.6883
  DiffEIC (s=50)	0.0422	0.1346	0.3501	0.7586
  OSCAR (s=1)	0.0507	0.1102	0.3487	0.7623
  PerCo (s=20)	0.0937	0.0993	0.2769	0.8081
  DiffEIC (s=50)	0.0945	0.0687	0.2275	0.8838
  OSCAR (s=1)	0.0937	0.0694	0.2285	0.8773

• Q3: Could the authors elaborate on the advantages of mapping bitrates to diffusion timesteps?

  The core idea of our approach is that for each bitrate, the quantization error can be well approximated as Gaussian noise at a specific diffusion timestep. By assigning the most appropriate timestep to each bitrate, OSCAR can more effectively leverage a pre-trained diffusion model for image compression. This design is validated by our extensive experiments, showing that OSCAR can support multiple bitrates using only one-step diffusion while achieving competitive or even superior performance compared to existing multi-step diffusion models that operate at a single bitrate.

  As for the bitrate control, most existing variable-bitrate methods rely on hyperparameters such as quality factor (e.g., BPG and VVC) and compression level (e.g., CTC). This typically requires users to tune these parameters to match a target bitrate. In contrast, our method allows users to specify the bitrate explicitly, thereby offering a more user-friendly interface for bitrate control.

• Q4: Can we combine this approach with existing variable-bitrate coding?

  As discussed in Q3, our bitrate-to-timestep mapping is specifically designed to model the quantized latent as a diffusion state, enabling better utilization of the pre-trained diffusion model for image compression. This design is fundamentally different from existing variable-bitrate methods that rely on quantization parameters to control compression levels. Therefore, these methods cannot be simply combined.

• Q5: The numerical value of the MACs and running time of the leading baselines should be provided.

  Thank you for the suggestion. We will include the number of MACs and the runtime (covering both encoding and decoding) in our paper.

  The results below are measured on images of size 1024$\times$1024.

  models	Encoding (ms)	Decoding (ms)	Total (ms)	MACs (T)
  MS-ILLM	208.77	314.72	523.49	1.05
  DiffEIC	478.70	21821.79	22300.49	304.87
  OSCAR	239.48	622.26	861.74	8.54

• Q6: Additional analysis on the RD-curves presented in Figure 7.

  We agree that RD-curves typically exhibit convex or concave trends. However, in our method, the bitrate is jointly determined by the downsampling ratio and codebook size. Specifically, 0.0937 and 0.1250 share the same downsampling ratio, as do 0.0313 and 0.0507. Bitrates with the same downsampling factor tend to show more similar distortion behaviors, leading to the observed patterns in our RD-curve. Appendix Figure 1 further illustrates this, bitrates with the same downsampling ratio also show similar cosine similarity. A similar trend is also observed in PerCo (Fig. 5(a)), which also uses a similar codebook-based design.

  RD-curves typically measure the distortion between reconstructed and original images. For no-reference IQA metrics such as MUSIQ and CLIP-IQA, the usual convex or concave trends may not apply, as these metrics are influenced by perceptual quality beyond simple image similarity. As a result, removing key modules may alter the curve’s shape more significantly. Therefore, while these metrics are valuable for comparing overall perceptual quality, they may be less suitable for analyzing RD curve trends.

  We confirm the consistency of our experimental setup and will release all code and models to ensure reproducibility.

• Q7: The authors should clarify the bitrate scalability of the model.

  Thank you for the helpful comment. We retrained OSCAR to include two lower bitrates and evaluated its performance in the extremely low-bitrate setting. The results show that OSCAR achieves competitive performance down to 0.0098 bpp, demonstrating its scalability across a broad bitrate range (0.0098 to 0.1250). These results will be included in the revised manuscript.

  The table below shows results on the DIV2K-val dataset. Numbers in parentheses indicate relative change compared to the original version of OSCAR.

  models	bpp	DISTS↓	LPIPS↓	MS-SSIM↑
  OSCAR	0.0313	0.1274 (↑0.95%)	0.3866 (↓0.44%)	0.7234 (0.33%↑)
  OSCAR	0.0507	0.1098 (↓0.36%)	0.3479 (↓0.23%)	0.7644 (0.28%↑)
  OSCAR	0.0937	0.0702 (↑1.15%)	0.2398 (↑0.55%)	0.8694 (↓0.22%)
  OSCAR	0.1250	0.0630 (↑0.96%)	0.2170 (↑0.60%)	0.8881 (↓0.28%)
  OSCAR	0.0234 (new)	0.1494	0.4265	0.6993
  PerCo (SD)	0.0234	0.2359	0.4433	0.6472
  OSCAR	0.0098 (new)	0.1820	0.4628	0.6556
  PerCo (SD)	0.0098	0.2870	0.4841	0.5889

We hope this response can help address your concerns. We believe our work can have an important impact on the field of image compression by offering a new perspective on leveraging pre-trained diffusion models for efficient one-step reconstruction across multiple bitrates. If you find our response satisfactory, we would be sincerely grateful if you could consider a higher score, as it would greatly support our efforts. Thank you once again for your consideration.

Thanks for your answer, but I think some important concerns are still not solved.

Firstly, I must point out some issues in your answer to Q2:

The original intent of my question was to know whether using only a single-step diffusion step might lead to greater semantic deviation compared to multi-steps. I know that the current perceptual metrics (e.g., LPIPS and DISTS) are limited, where higher scores may not mean better fidelity. That's also the reason that why I ask for the visual comparison with DiffEIC (The visual comparison could visually demonstrate the difference in fidelity). I observed that you show the performance of Perco on high-resolution DIV-2k, however, that is not valuable. Perco performs much worse on high resolution images than low-resolution ones due to its design. They target at low-resolution in their paper.

models bpp DISTS↓ LPIPS↓ MS-SSIM↑

DiffEIC (s=50) 0.0422 0.1346 0.3501 0.7586

OSCAR (s=1) 0.0507 0.1102 0.3487 0.7623

I think for this rate point, you could not claim OSCAR achieves performance competitive with DiffEIC (DiffEiC use less bits but achieve much better DISTS) and significantly surpasses PerCo (You should compare the low-resolution datasets) in both perceptual and pixel-level fidelity.

About Q6: Abnormal RD-Curves, which I think is the biggest concern.

I think the simplest analysis is to test more rate points to show the RD-Curve with more details(Since your model supports variable bitrates). Non-convex or Non-concave curves seems to show that your solution may show performance drops (e.g., PSNR degradation) at certain bitrates, which this indicates a fundamental instability in the model’s rate control. It is necessary for variable bitrate models: As a variable bitrate model, it must ensure smooth performance transitions when the bitrate changes continuously. If performance in certain intervals is significantly worse than that of adjacent bitrate points (even at the same downsampling ratio), this should be attributed to model design (such as the codebook allocation strategy or downsampling-reconstruction coupling).

Finally, about Q4: I think the technique of existing variable-bitrate coding, which usually modulate the latent with a scaler does not conflict with your mechanism that use different diffusion steps. It is just a suggestion that you may could have a try in the future, but not a necessity.

Thanks for your valuable comment, nice suggestions, and for acknowledging our writing and empirical performance. Your questions and suggestions are instrumental in further strengthening our paper. Below, we respond to your specific comments.

• Q1: The mathematical explanation and the motivation could be discussed in more detail.

  Thank you for this insightful comment. We acknowledge that the core of OSCAR is based on an empirical assumption. While this assumption is not derived from a rigorous mathematical formulation, it is strongly supported by experimental evidence.

  The motivation behind OSCAR is as follows: We begin by recalling the standard formulation of the forward diffusion process: $z_t=\sqrt{\alpha_t} z_0+\sqrt{1-\alpha_t}\epsilon$, where $\epsilon \sim \mathcal{N}(0,1)$. Since we aim to use the diffusion formulation to reconstruct the clean latent $z_0$ from the quantized latent $\tilde{z}$, we naturally ask the question, can $\tilde{z}$ also be expressed in the form $\tilde{z}=\sqrt{\alpha} z_0+\sqrt{1-\alpha}\epsilon$? We test different values of $\alpha$ and visualize distribution of the quantization error, computed as $\epsilon=(\tilde{z}-\sqrt{\alpha} z_0)/ (\sqrt{1-\alpha})$. Interestingly, we find that after the first-stage alignment, it is possible to select a suitable $\alpha$ such that the resulting quantization error closely approximates a Gaussian distribution, as illustrated in Figure 4. For a fixed bitrate, we find that the selected $\alpha$ values are similar across different images. This motivates us to statistically determine the most appropriate $\alpha$ for each bitrate. Specifically, we compute the average cosine similarity across samples in our training set and select the pseudo timestep $t$ whose theoretical cosine similarity $\sqrt{\bar{\alpha}_t}$ best matches the observed value. This mapping procedure ensures a stable and interpretable assignment of bitrates to diffusion timesteps.

  In the introduction section L45–59, we outline the high-level motivation. In Section 3.4, we provide detailed formulations of how we map the bitrates to pseudo diffusion timesteps. Appendix C provides a mathematical justification for the validity of using cosine similarity to establish the mapping. We will consolidate these discussions in the main text to provide a more coherent explanation.

• Q2: Whether the method's effectiveness stems from inductive biases rather than the use of diffusion.

  We would like to clarify the role of diffusion in our method. In Section 3.1, we provide the mathematical formulation of one-step diffusion. The only adaptation we introduce in OSCAR is treating the quantized latent as the noisy input to the denoising process—no changes are made to the core diffusion architecture or its operations. This design remains fully consistent with the one-step diffusion paradigm.

  To demonstrate the role of diffusion, we removed the time module in the diffusion backbone and let the diffusion backbone predict the clean latent directly (thus making the backbone a pure denoiser). We trained this denoiser with the same experimental setting as OSCAR. Below, we report the performance on the DIV2K-val dataset.

  models	bpp	DISTS↓	LPIPS↓	MS-SSIM↑
  OSCAR	0.0313	0.1262	0.3883	0.7210
  Denoiser	0.0313	0.1687	0.4259	0.6761
  OSCAR	0.0507	0.1102	0.3487	0.7623
  Denoiser	0.0507	0.1479	0.3726	0.7254
  OSCAR	0.0937	0.0694	0.2385	0.8713
  Denoiser	0.0937	0.0933	0.2697	0.8224
  OSCAR	0.1250	0.0617	0.2157	0.8906
  Denoiser	0.1250	0.0887	0.2514	0.8504

  The diffusion formulation shows a clear advantage over the denoiser. We attribute this to our design strictly following the one-step diffusion framework, thus allowing us to fully leverage pre-trained diffusion backbones.

• Q3: Comparison with more recent baselines.

  Thank you for the advice. We will supplement these comparisons in our paper, as well as a discussion of similar works in the related work section.

  1. Comparison with IPIC.

     The table below presents results on the Kodak dataset.

     models	bpp	DISTS↓	LPIPS↓	MS-SSIM↑
     IPIC (s=1000)	0.0721	0.1674	0.3885	0.8221
     OSCAR (s=1)	0.0507	0.1156	0.3360	0.7610
     IPIC (s=1000)	0.1200	0.1640	0.2811	0.8447
     OSCAR (s=1)	0.1250	0.0766	0.2340	0.8585

     The results show that OSCAR performs better than IPIC on DISTS and LPIPS, while IPIC shows better MS-SSIM at low bitrates. This could be attributed to the generative nature of the diffusion backbone, which prioritizes perceptual quality over pixel-wise reconstruction.

  2. Comparison with DDCM.

     In our method, we use LPIPS-AlexNet as a training loss, and LPIPS-VGG as the evaluation metric to ensure a fair comparison. Here we follow DDCM to use LPIPS-AlexNet (denoted as LPIPS*), PSNR, and FID for evaluation, and use the DIV2K-val dataset with images cropped to 768$\times$768.

     model	bpp	PSNR↑	LPIPS*↓	FID↓
     DDCM (s=1000)	0.0420	21.1670	0.2010	18.3300
     OSCAR (s=1)	0.0313	21.3304	0.1958	31.8673
     DDCM (s=1000)	0.0660	21.7800	0.1720	16.3960
     OSCAR (s=1)	0.0507	21.9478	0.1554	28.9917
     DDCM (s=1000)	0.1370	22.7070	0.1380	14.6620
     OSCAR (s=1)	0.1250	23.4529	0.1097	15.7407

     Compared to the 1000-step DDCM, our one-step OSCAR performs better on reference-based metrics (PSNR and LPIPS*), suggesting higher fidelity to the original images. While DDCM achieves lower FID scores, the gap narrows as the bitrate increases.

• Q4: The method relies heavily on a large pre-trained diffusion model.

  We appreciate your acknowledgment that, despite the reliance on a pre-trained diffusion model, our one-step diffusion design takes an important step to reduce the time burden. Here we would like to supplement that leveraging pre-trained diffusion models has become a common practice in recent works—PerCo, DiffEIC, and DDCM all rely on them. Compared to these multi-step diffusion codecs, our one-step approach not only greatly reduces computational overhead but also achieves competitive or even better performance.

We hope this response can help address your concerns. We believe our work can have an important impact on the field of image compression by offering a new perspective on leveraging pre-trained diffusion models for efficient one-step reconstruction across multiple bitrates. If you find our response satisfactory, we would be sincerely grateful if you could consider a higher score, as it would greatly support our efforts. Thank you once again for your consideration.

Thank you for your valuable comments and insightful suggestions, as well as for acknowledging our novelty, motivation, and convincing evaluations. Below, we respond to your comments.

• Q1: The authors should discuss the generalization ability to unseen or non-predefined bitrates.

  Thanks for the suggestion. Since our method employs a learnable hyper-encoder for compression, it does not support unseen bitrates out of the box. However, we conducted additional experiments to demonstrate OSCAR’s ability to quickly adapt to new bitrates with minimal computational overhead. We adopted a simple training strategy: half of the steps are allocated to the new bitrate, and half to the original groups. Extending to a new bitrate takes only 15K steps—under 4 hours on 4×A6000 GPUs. We attribute this efficiency to the fact that the diffusion model has already encountered structurally similar latent distributions during initial training, enabling fast and stable adaptation to nearby bitrates.

  The tables below report results on the DIV2K-val dataset, where we progressively adapt the model to new bitrates one at a time. Numbers in parentheses indicate the performance change compared to the model before adaptation to the new bitrate.

  bpp	DISTS↓	LPIPS↓	MS-SSIM↑
  0.0313	0.1262	0.3883	0.7210
  0.0507	0.1102	0.3487	0.7623
  0.0937	0.0694	0.2385	0.8713
  0.1250	0.0617	0.2157	0.8906

  bpp	DISTS↓	LPIPS↓	MS-SSIM↑
  0.0313	0.1306 (↑3.49%)	0.3927 (↑1.13%)	0.7161 (↓0.68%)
  0.0507	0.1141 (↑3.54%)	0.3523 (↑1.04%)	0.7578 (↓0.59%)
  0.0937	0.0728 (↑4.90%)	0.2427 (↑1.77%)	0.8595 (↓1.35%)
  0.1250	0.0648 (↑4.93%)	0.2198 (↑1.91%)	0.8766 (↓1.57%)
  0.0430 (new)	0.1198	0.3670	0.7402

  bpp	DISTS↓	LPIPS↓	MS-SSIM↑
  0.0313	0.1360 (↑4.19%)	0.4006 (↑2.03%)	0.7031 (↓1.81%)
  0.0507	0.1185 (↑3.87%)	0.3589 (↑1.87%)	0.7459 (↓1.57%)
  0.0937	0.0752 (↑3.34%)	0.2457 (↑1.25%)	0.8506 (↓1.03%)
  0.1250	0.0672 (↑3.75%)	0.2228 (↑1.38%)	0.8664 (↓1.16%)
  0.0430	0.1246 (↑4.01%)	0.3742 (↑1.98%)	0.7282 (↓1.62%)
  0.0781 (new)	0.0918	0.2766	0.8412

• Q2: The assumption of the method might not hold at low bitrates.

  We would like to clarify that although the cosine similarity between quantized and clean latents naturally decreases at lower bitrates, the assumption that the quantization error follows additive Gaussian noise still holds, as long as the cosine similarity remains stable during the second-stage training.

  To further address your concern, we retrained OSCAR from scratch to include two additional lower bitrates using the same experimental settings, except that we increased the training iterations to 150k. The results show that our method remains effective at these lower rates. We also visualize the quantization error using the same method as in Figure 4, and the distribution continues to closely follow a Gaussian. These results will be included in the revised manuscript.

  The tables below report results on the DIV2K-val dataset. Numbers in parentheses indicate the performance change relative to our original model.

  models	bpp	DISTS↓	LPIPS↓	MS-SSIM↑
  OSCAR	0.0313	0.1274 (↑0.95%)	0.3866 (↓0.44%)	0.7234 (0.33%↑)
  OSCAR	0.0507	0.1098 (↓0.36%)	0.3479 (↓0.23%)	0.7644 (0.28%↑)
  OSCAR	0.0937	0.0702 (↑1.15%)	0.2398 (↑0.55%)	0.8694 (↓0.22%)
  OSCAR	0.1250	0.0630 (↑0.96%)	0.2170 (↑0.60%)	0.8881 (↓0.28%)
  OSCAR	0.0234 (new)	0.1494	0.4265	0.6993
  PerCo (SD)	0.0234	0.2359	0.4433	0.6472
  OSCAR	0.0098 (new)	0.1820	0.4628	0.6556
  PerCo (SD)	0.0098	0.2870	0.4841	0.5889

• Q3: Authors should provide more theoretical and cross-model generalization analysis.

  Thanks for the valuable advice. We agree that OSCAR is based on an empirical assumption, which we validate through extensive experiments, rather than deriving from a strictly theoretical framework. While establishing a formal theoretical mapping is certainly an interesting direction, it remains challenging and still has a long way to go. In this context, we propose an alignment-based approach, where we empirically model the quantization error as Gaussian noise, based on the observed cosine similarity between the quantized and original latents. This assumption is strongly supported by our experiments. As shown in Figure 4, the quantization error closely follows a Gaussian distribution. We will include additional visualizations of quantization error at different bitrates in our paper. We believe that our perspective, supported by empirical findings, will provide valuable insights for future research.

  To assess cross-model generalization, we replaced the UNet-based SD-2.1 with the Transformer-based SD-3 Medium and conducted experiments accordingly. We report the results on the DIV2K-val dataset. The results show that our method remains effective under this different backbone, demonstrating cross-model generalization.

  models	bpp	DISTS↓	LPIPS↓	MS-SSIM↑
  OSCAR (UNet)	0.0313	0.1262	0.3883	0.7210
  OSCAR (DiT)	0.0313	0.1076	0.3579	0.7532
  OSCAR (UNet)	0.0937	0.0694	0.2385	0.8713
  OSCAR (DiT)	0.0937	0.0657	0.2124	0.9003

• Q4: One hyper-encoder is trained for each bitrate, which might limit the generalization to fine-grained bitrate settings.

  We agree that generalizing to fine-grained bitrate settings remains a challenge, especially when aiming for both flexibility and high performance. However, as shown in Appendix Figure 1, despite using separate hyper-encoders, the quantized latents of nearby bitrates remain highly similar, mirroring the behavior of variable-bitrate methods. We believe this latent consistency suggests that OSCAR has the potential for future improvements.

• Q5: Would it be possible to allow different hyper-encoders to share modules?

  We would like to clarify that each hyper-encoder is tied to a specific bitrate because each uses different internal settings, like the downsampling ratio and codebook size, to control the compression level. The module requires no complex design and is very lightweight ($\approx$0.5% of the parameters compared to the diffusion backbone). We found that having a separate lightweight hyper-encoder per bitrate works well and keeps things simple. Nevertheless, an interesting future direction is to share a single hyper-encoder across bitrates that use the same downsampling ratio but differ in codebook size. This would allow multiple bitrates to be paired with one hyper-encoder.

• Q6: How can this method be extended to continuous bitrate adaptation in future work?

  Thank you very much for the inspiring question. One important current limitation lies in the total training budget. Under constrained resources (e.g., 4×A6000 GPUs), it is challenging for a single model to fully optimize across many bitrates without sacrificing performance. A straightforward solution is to scale up training, e.g., using larger batch sizes and more iterations to allow the model to better generalize across bitrate levels.

  Furthermore, our results (Appendix Figure 1) show that quantization errors tend to be similar when the downsampling ratios are the same. This suggests a promising future direction: we could train a model on several anchor bitrates with different downsampling ratios, and then generalize to additional bitrate levels, which share the same downsampling ratio as the anchor bitrate but with different codebook sizes. In fact, as discussed in Q1, our simple training strategy already allows fast adaptation to new bitrates with minimal cost, leveraging well-trained anchor bitrates. We believe that designing efficient and scalable mechanisms for bitrate extension is a very promising direction.

We hope this response can help address your concerns. We believe our work can have an important impact on the field of image compression by offering a new perspective on leveraging pre-trained diffusion models for efficient one-step reconstruction across multiple bitrates. If you find our response satisfactory, we would be sincerely grateful if you could consider a higher score, as it would greatly support our efforts. Thank you once again for your consideration.

Thank you for your acknowledgement. We hope our response has satisfactorily addressed your concerns, and we will incorporate these important empirical results and discussions into the revised version. May we kindly ask if there are any remaining concerns? We would be more than happy to address them. Thank you once again for your time and thoughtful consideration.