MoRIC: A Modular Region-based Implicit Codec for Image Compression

We thank the reviewer for their valuable feedback. Here is our point-to-point response:

Response to each weakness:

• W1. We currently use SAM2 for region partitioning, which inherently supports automatic segmentation and can be extended with other advanced strategies such as saliency-based or prompt-guided segmentation for various applications. For alternative partition methods, we refer the reviewer to our response to [3Cij] (W1. and Q3.)

• W2. The reviewer is correct. We also want to note that while MoRIC falls short of the MLIC++ in RD performance for standard image compression, it achieves over 100× lower decoding complexity. The superior performance of MLIC++ in the low-bpp regime for single-object scenarios can be attributed to its natural advantage in this setting. In contrast, overfitted codecs face inherent limitations due to the dominance of parameter cost in this regime.

  In response to this comment, we have added the following clarification in the revised manuscript: "Although MoRIC shows significant RD gains, it still shares the common low-bpp performance bottleneck with other overfitted codecs due to dominant parameter cost, a challenge that Bayesian INR parameter [1] coding may help address."

  [1] He, Jiajun, et al. "RECOMBINER: Robust and Enhanced Compression with Bayesian Implicit Neural Representations."

• W3. We thank the reviewers for pointing out this issue. While MoRIC has lower theoretical decoding complexity, its observed latency remains slightly high due to the lack of engineering-level optimizations. We want to note that real-world latency is affected by many uncontrollable factors in a lab setting and can be significantly reduced through various optimization techniques. For example, [2] reports 100 ms decoding time for COOL-CHIC using a C API and binary arithmetic coding, whereas our implementation runs without such optimizations. Notably, MoRIC is implemented with PyTorch serial inference across all modules for research purposes, where substantial speedups are expected with appropriate engineering optimizations.

  To improve clarity over latency, we have added the following clarification to the revised manuscript: "We want to note that all reported MoRIC results are based on unoptimized research code. Notably, our decoding modules are now executed serially during latency evaluation; a simple parallel execution can substantially reduce runtime. We also expect further improvements through advanced engineering optimizations such as C-API integration and binary arithmetic coding [2]."

  [2] Blard, Théophile, et al. "Overfitted image coding at reduced complexity."

• W4. Due to computational constraints, initial CLIC2020 results were provided in our supplementary materials (Table 1). We have since conducted comprehensive experiments with more RD values and adaptive architectures. (see Table 1.1, 1.2, and 1.4 in our response to [Ek8z].)

Response to each question:

• Q1. The reviewer is correct. We will revise the manuscript to include the following additional sentence: "C* coding approximates object contours using polygonal chains. Starting from an initial point, each contour segment is encoded by a fixed step length and a quantized offset angle relative to the previous direction. A move-to-front transform is applied to the angle sequence, alongside adaptive mechanisms such as global resets, offset adaptation, and contour dilation, to enhance compression efficiency." Experiments about the chain-coding are also provided in Table 3.3 in our response to [BiKZ].

• Q2. The contour boundaries remain barely visible, even at low bitrates such as 0.05 bpp. This robustness stems from two key design choices in MoRIC:

  (a). Design of the global modulation network (GMN): MoRIC integrates global-local context through its GMN and progressive modulation, effectively suppressing artifacts. Similar strategies have been shown to reduce artifacts in INR literature [3–4].

  (b). Region-wise fitting strategy: Instead of fitting the image in a patch-wise manner, MoRIC performs region-wise fitting, enabled from C* coding, enhancing INR representation while reducing artifacts (as also observed in [3]).

  Following the suggestion, we will add visualizations to show artifact suppression, especially at low BPPs. To further quantify this in the discussion period, we include an ablation study on GMN and region-wise fitting (exp-1 and exp-2) for artifacts analysis and quality assessments, reporting PSNR, MS-SSIM, PSNR-edge/PSNR-smooth edge (two metrics from [4] that assess artifacts via Canny edge detection with disk dilation). For exp-1, we compare MoRIC with a variant that removes the GMN component, denoted as MoRIC (w/o GMN). Exp-2 replaces the target region of MoRIC with a rectangular patch to cover the same target, denoted as MoRIC (rect), where contour coding is omitted. Detailed performance with BD-rate increase for each variant of MoRIC is presented in Table 4.1.

Table 4.1-a Ablation study for MoRIC (w/o GMN) for artifacts analysis on Kodak.

Table 4.1-b Ablation study for MoRIC (rect) for artifacts analysis on Kodak dataset.

From Exp-1, we conclude the importance of GMN: without it, ringing artifacts and inconsistent texture scale can be observed. From Exp-2, we can observe noticeable stitching artifacts in MoRIC (rect) with reduced RD and perceptual performance, highlighting the effectiveness of the C* coding. To further assess MoRIC’s perceptual quality, we report MS-SSIM results under MS-SSIM-trained settings (Table 4.2). MoRIC achieves impressive performance, empowered by the above two design choices, closely matching advanced codecs like ELIC. Both Table 4.1 and Table 4.2 results will be incorporated into the revised manuscript.

[3] Ashkenazi, Maor, and Eran Treister. *"Towards croppable implicit neural representations."

[4] Feng, Haoan, et al. *"SASNet: Spatially-Adaptive Sinusoidal Neural Networks."

Table 4.2: Detailed MS-SSIM / bpp performance on Kodak dataset.

• Q3. Following the reviewer’s suggestion, we conducted additional experiments by training MoRIC using masks with varying degrees of over- and under-segmentation errors.

  Specifically, we simulate small and substantial segmentation errors via mask dilation and erosion, with IoU values of 14.22%/34.18% (under-segmentation) and 19.85%/48.58% (over-segmentation). Results are shown in Table 4.3.

  We observe that MoRIC is robust to segmentation imperfections, especially when the segmentation IoU is below 20%, with performance degrading gradually as segmentation quality decreases. However, for both under- and over-segmentation, more significant performance drops begin to appear when the IoU approaches 30–40%. Additionally, substantial segmentation errors can introduce encoding instability, occasionally leading to suboptimal RD performance.

  While MoRIC does not currently optimize region selection, we believe that better segmentation strategies could yield substantial improvements by leveraging this trade-off.

Table 4.3 Ablation study over segmentation error for MoRIC on Kodak dataset.

We thank the reviewer for the valuable feedback. In response, we will: (a) include an ablation study on the robustness of partition strategy in MoRIC; (b) introduce an automatic partitioning algorithm with adaptive region count; (c) add experiments demonstrating MoRIC’s encoding advantages over other overfitted codecs; and (d) provide a detailed analysis of the trade-offs between region count, complexity, and bitrate.

Next, we address the commented weakness and the question point by point.

• W1. We thank the reviewer for highlighting this important point. While our current implementation uses SAM2 for region selection, we agree that fully automated partitioning is essential for practical deployment.

  To this end, we note that MoRIC supports automatic region partitioning via connected component analysis on SAM2 outputs, enabling automatic selection of the top-k object regions. It is also compatible with task-adaptive strategies, including prompt-guided segmentation from SAM2, as well as alternatives like DINOv2 with DeepLabv3 or EML-Net for saliency-based partitioning, allowing enhanced adaptability and partition quality across diverse applications.

  It is worth noting that the performance of MoRIC is robust to segmentation imperfections, thanks to its advanced PCM design and region-wise INR fitting, as demonstrated in our response to Reviewer [1oDR] (Table 4.3).

  We will add the following clarification to the revised manuscript:
  “MoRIC is robust to segmentation quality, supports automatic region partitioning via SAM2, and can incorporate task-adaptive strategies such as prompt-guided or saliency-aware segmentation to enhance partition quality and compression performance.”

• W2. High encoding complexity is a common phenomenon across all overfitted coding paradigms due to their instance-specific optimization, and is not unique to MoRIC. Potential solutions include meta-learning [1] and engineering optimizations [2], which have been shown to accelerate encoding by 10×–100×, reducing Kodak sample encoding time to around 20 seconds.

  In this context, we would like to highlight another contribution of our work (previously discussed in Supplementary Section 3.1): MoRIC mitigates encoding complexity to some extent via a region-wise fitting strategy with shared modulation, which accelerates convergence and improves practical efficiency.

  This design significantly reduces the number of required encoding steps compared to prior overfitted codecs, achieving a 2–3× algorithmic speedup (see Table 3.2 in our response to Reviewer [BiKZ]). These convergence gains, combined with potential engineering improvements, represent meaningful progress toward practical, low-latency overfitted codecs.

  [1] Dupont, Emilien, et al. "COIN++: Neural Compression Across Modalities." Transactions on Machine Learning Research.

  [2] Blard, Théophile, et al. "Overfitted image coding at reduced complexity." 2024 EUSIPCO.

Answers to questions:

• Q1. Following reviewer's suggestion, we provide a detailed analysis of the trade-off between region count (N) and coding complexity. As shown in Table 2.1, the breakdown of coding cost is reported for various values of N (1–3). (Considering various existing implementations of entropy coding schemes and for fair comparison with [3], we exclude range coding time from our measurements, which is also a component of every classical codec.)

  [3] Kim, Hyunjik, et al. "C3: High-performance and low-complexity neural compression from a single image or video." CVPR. 2024.

Table 2.1 Encoding (1000 steps) and decoding cost breakdown of MoRIC on Kodak dataset for different numbers of regions.

SAM2 partition time: 533–830 ms (GPU); C* encoding: 119.58 ms (CPU); C* decoding: 12.86 ms (CPU)

From the table, we observe that increasing the number of regions leads only to a slight rise in overall complexity. Specifically, each additional local INR adds 33 MACs/pixel from LSNs and 264 MACs/pixel from GMN in the forward pass. The added cost primarily arises from extra operations in the shared modulator to accommodate more local INRs (see Fig. 5 in our manuscript).

We also recognize the importance of real-world latency and thereby include a breakdown of actual coding time across regions to provide practical insights. Here the time for the partition and contour coding is also provided. From our experiments, ARM exhibits consistent time costs across different region settings. Each additional LSN adds approximately $8$–$9$ ms to decoding time and $2$–$4$ seconds for 1000 encoding steps. A similar trend is observed for GMN, which dominates the total coding time and increases significantly as the number of regions grows, consistent with its complexity analysis. Note that all reported results are based on the same unoptimized code, with decoding executed serially. Significant speedups are expected for MoRIC with proper engineering optimizations.

Overall, MoRIC demonstrates strong scalability with respect to region count, as the added complexity and latency remain acceptable relative to its strong rate-distortion performance. We will include the above discussion and Table 2.1 in the revised manuscript to provide a more comprehensive analysis of complexity and coding time across different region numbers.

• Q2. We repeat the experiments from Fig. 9 in the manuscript and provide a detailed breakdown of the bitrate share between contours, network parameters, and latent for different rate regimes and region numbers $N$ (see Table 2.2).

  From the table, we can observe that C* coding can compress contours with minimal bitrate overhead, remaining as low as 1.89% for $N=2$ and increasing only slightly with more regions (up to 3.20% for $N=3$). As bitrate increases, the relative impact of contour cost becomes negligible, allowing MoRIC to achieve competitive RD performance with finer region partitioning efficiently.

  We will include the above discussion and Table 2.2 in the revised manuscript to provide a more comprehensive analysis of bitrate distribution across different compression rates and region numbers.

Table 2.2 Bit rate breakdown for MoRIC across region numbers on Kodak dataset.

• Q3. There exist adaptive strategies for determining the number of regions and corresponding partitions for MoRIC. In general, images with substantial regional variation benefit from using more local INRs, which enables faster fitting and improved compression performance.

  This observation suggests that the problem of getting the adaptive region number can be transformed into an analysis of the complexity of an image’s regional distribution. While a full investigation of how region distribution variation impacts partitioning and corresponding RD performance is beyond the scope of this paper, we consider this a valuable suggestion and a promising direction for future development of implicit codecs. To this end, we outline two practical solutions:

  (a) Feature-driven region partitioning: Leverage a strong backbone such as DINOv2 for robust region-feature extraction and semantic grouping, followed by SAM2 to perform region segmentation guided by these cues. This combination enables adaptive region partitioning that captures region-wise semantic structure, where more complex images naturally result in more regions guided by semantic grouping.

  (b) Saliency-guided region partitioning: A learning-based saliency model such as EML-Net can be used to guide partitioning by grouping visually important or semantically similar areas. Subsequent posterior analysis and filtering can further refine the selection to balance coding complexity and rate-distortion performance.

  These approaches demonstrate MoRIC’s versatility and strong potential for integration with emerging methods and diverse application scenarios.

Dear reviewer,

Thank you for the review.

As the discussion phase is ending soon, can you please go over the authors' response and acknowledge it?

Best, AC

We thank the reviewer for their valuable feedback. In response, we will:
(a) include a discussion on practical applications for scalable layered compression;
(b) discuss automatic partitioning alternatives;
(c) add experiments analyzing the trade-off between performance and coding cost;
(d) add experiments to validate the efficiency of C* coding; and
(e) include additional experiments on the CLIC2020 dataset.

Next, we address the identified weakness and the questions point by point.

• W1. To highlight scalable layered compression applications of MoRIC, we have revised the manuscript to include the following section:

  "MoRIC supports progressive rate control with semantic consistency, enabling several practical applications:

  (i) Goal-oriented progressive compression: Suited for scenarios like surveillance and edge AI, where full-resolution reconstruction is not always needed. Semantic outlines are sent first, with key regions refined on demand, saving bandwidth while preserving critical information.

  (ii) Distributed compression: In multi-agent settings (e.g., robotics, NeRF), agents transmit only observed regions; others are inferred or selectively refined in a MoRIC fashion.

  (iii) Interactive content delivery (e.g., VR, AR, games): Decoding adapts online to user attention, delivering high-quality reconstruction in focus areas, while other regions are inferred and progressively refined as attention shifts.

  (iv) Video streaming/preview: Coarse frames with accurate key regions and semantically consistent background are transmitted first, followed by progressive refinement for a seamless viewing experience."

  To further illustrate the benefits of partial decoding, we report detailed computation and bit cost savings for a surveillance scenario in Table 3.1. For experimental details, we refer to Fig. 5 in our supplementary materials for a demo illustration.

  MoRIC achieves up to 73.56% and 37.31% bitrate savings compared to VTM and MLIC++, respectively. It also provides substantial computational savings, thanks to its low decoding complexity, where the complexity of VVC is estimated to be comparable on standard hardware.

Table 3.1. Bitrate and computational savings enabled by partial decoding in MoRIC

• W2. We currently use SAM2 for region partitioning, which inherently supports automatic segmentation and can be extended with other advanced strategies such as saliency-based or prompt-guided segmentation for various compression scenarios. For alternative partition methods, we refer to our response to reviewer [3Cij] (W1. and Q3). For robustness to segmentation quality, we refer to Table 4.3.

• W3. In response to the comment, we include an experiment evaluating the performance–coding cost trade-off. We train MoRIC (lossy-fixed) and CoolChic-V4 on the Kodak dataset under encoding steps from 1k to 80k. The resulting BD-rates over VTM-19.1 are shown in Table 3.2. MoRIC achieves comparable RD performance with 2–3× fewer coding steps, attributed to its region-wise coding strategy, similar to the observations in [1,2].

  Consequently, MoRIC can instead potentially accelerate convergence rather than slowing encoding. As shown in Table 3 and Fig. 3 in the supplementary, increasing the number of regions (from 1 to 3) can further improve convergence through more effective INR specialization.

  Previously, encoding cost was typically addressed via engineering optimizations for an overfitted codec. In this context, a 2–3× algorithmic acceleration represents a meaningful step toward practical, low-latency compression. While our results reflect theoretical gains in convergence and complexity, actual latency also depends on deployment factors such as memory and hardware, where further improvements are expected.

  [1] "Partition speeds up learning implicit neural representations based on exponential-increase hypothesis", 2023.
  [2] Coordx: Accelerating implicit neural representation with a split MLP architecture, 2022.

  In addition to gains in performance, acceleration, and convergence, MoRIC also enables applications such as efficient target-area compression and layered compression. Together, these results demonstrate the effectiveness of the MoRIC design.

Table 3.2. BD-rate over VTM-19.1 for different schemes across various encoding steps

Answers to questions:

• Q1. For the evaluation of the single-object compression task, only the target region is considered, with masked pixels excluded from computation. We will clarify this in the revised manuscript.

• Q2. We explored multiple design choices and adopted the best-performing configurations. By default, each baseline codec operates on the smallest rectangular region enclosing the target object, with masked pixels within this region jointly learned. For INR-based methods, we also evaluate variants combined with our proposed C* coding algorithm for ablation studies on the architectural innovations. For more details of the baseline models for single-object cases, we refer to Section E.2 of the appendix (Lines 638–655).

Table 3.3. Contour coding efficiency on Kodak dataset (kodim23), evaluated by error rate / bit cost

• Q3. In response, we provide both theoretical analysis and experiments (Table 3.3) to support and verify our findings.

  Whether contour coding is more efficient than mask-based encoding depends on the applications. When the goal is to only localize and compress object boundaries, chain coding can offer significant advantages (see Table 3.3 as an example). Mask-based approaches typically require one bit per pixel across the entire image, whereas contour representations involve a significantly smaller set of pixels. Instead, a basic chain coding scheme can encode directional steps (e.g., 8 directions) using just 3 bits per step. More advanced methods that incorporate geometric cues and relative motion (such as VCC, 3OT, or NAD) can further enhance compression efficiency. These techniques are all included as open-source resources in our supplementary materials.

  As shown in Table 3.3, our C* coding method achieves lossless compression (error rate 0%) with a comparable bit cost to VTM’s mask coding (with an error rate of 6.62%). Moreover, by adjusting the step length, C* coding can further reduce the bitrate with only minimal degradation in coding accuracy.

  However, chain coding may be less efficient in scenarios with very noisy or highly irregular contours, where frequent direction changes and large curvature variations reduce coding efficiency. In such cases, mask-based approaches, especially those leveraging learned priors, can offer greater robustness.

  In our setting, contour coding aligns naturally with the design of overfitted codecs and can be integrated seamlessly into the framework.

• Q4. We have now reported a breakdown of the coding cost (Table 2.1 of our response to Reviewer [3Cij]), including all components except range coding time, to ensure a fair comparison with C3. We will clarify this point in the revised manuscript.

• Q5. We use the DAVIS dataset because it provides region-level ground truth, which aligns well with our experimental setup. Instead, standard compression datasets typically lack such annotations. Within our MoRIC framework, all region partitions are generated using SAM, and DAVIS allows us to validate our method on high-resolution data with reliable region labels. Additionally, we now include new experiments on the CLIC2020 dataset to further assess MoRIC’s performance (see Tables 1.1, 1.2, and 1.4 in response to Reviewer [Ek8z]).

  For small-object cases, baseline methods also benefit from reduced fitting regions (a smaller rectangular), while the relative parameter cost of the INR in MoRIC becomes more pronounced. Intuitively, MoRIC offers greater advantages in scenarios with complex object contours, where its explicit contour-based coding provides more substantial benefits.

• Q6. For the ablation study in Table 7, modules are directly removed to assess their individual contributions, without adjusting for complexity or parameter count. We want to note that ensuring strict fairness in parameter count for overfitted codec is challenging, as additional parameters may introduce overhead and can sometimes even degrade RD performance. Here, we focus on validating the impact of the C coding scheme, PCM modulation, and LSN design, to assess whether introducing these mechanisms leads to improved RD performance.

We thank the reviewer for the valuable feedback. In response, we will: (a) add descriptions of our partitioning strategy and include ablation studies on the impact of segmentation quality; (b) provide a detailed breakdown of coding cost across various regions, along with corresponding theoretical complexity analysis; (c) include additional experiments on general natural scenes (CLIC2020 dataset); and (d) incorporate more baseline comparisons in the revised manuscript.

Next, we address the commented weakness and the questions point by point.

• W1. The current MoRIC region partitioning is guided by semantics and based on SAM2 [1]. The reviewer is correct that partitioning may result in unreasonable segmentation. However, MoRIC does not rely on highly accurate segmentation, thanks to the robustness of its GMN design and INR fitting capability. In fact, while lossy contour coding may cause minor over-segmentation, the resulting bit savings without significantly sacrificing distortion often lead to slightly improved rate-distortion trade-offs.

  To evaluate this, we conducted experiments assessing the impact of segmentation quality under both over- and under-segmentation scenarios. The results are presented in Table 4.3 (in our response to Reviewer [1oDR]).

  The partitioning time of MoRIC is around 533-830 ms (3090GPU) for Kodak samples, which is reported in Table 2.1 (in our response to Reviewer [3Cij]) as part of the coding cost. Since overfitted codecs are typically encoded offline, this one-time partitioning will have minimal impact on practical deployment.

  In summary, MoRIC favors semantic partitioning over pixel-level precision due to its robustness to contour errors. Experimental results show that this design is both efficient and effective, while also maintaining strong perceptual quality, as demonstrated by the MS-SSIM experiments in Table 4.2 (in our response to Reviewer [1oDR]).

  [1] Ravi, Nikhila, et al. "SAM 2: Segment Anything in Images and Videos." ICLR 2025.

• W2. In addition to Table 4 in the manuscript, we will include a module-wise breakdown of coding time and a theoretical complexity analysis (which is detailed in Table 2.1 from our response to Reviewer [3Cij] due to space limitations) into our manuscript. Specifically, the backward pass complexity is approximated to twice the MACs of the forward pass, following [2].

  [2] Baydin, Atilim Gunes, et al. "Automatic differentiation in machine learning: a survey." Journal of Machine Learning Research (2018).

• W3. (a) Performance concerns: We would like to highlight that MoRIC sets a new state-of-the-art among overfitted codecs, outperforming VTM (H.266) on natural scene datasets such as Kodak and CLIC2020, a strength also noted by Reviewers [3Cij], [BiKZ] and [1oDR]. While MoRIC falls short of leading neural codecs like MLIC++ in RD performance, it achieves over 100× lower decoding complexity.

  To further demonstrate its advantages on general natural scenes, we conducted additional experiments on the CLIC2020 dataset (as shown in Tables 1.1, 1.2, and 1.4), where even more pronounced gains were observed (approaching MLIC and outperforming ELIC on CLIC2020). Additionally, we also provide MS-SSIM results for MoRIC (see Table 4.2 in our response to Reviewer [1oDR]), where MoRIC approaches ELIC on Kodak. Taken together, these results offer a comprehensive evaluation and underscore the promise of MoRIC, with evidence that is both strong and well-supported.

  Beyond the RD experimental gains, MoRIC introduces broader contributions, including a modulated local-region fitting strategy, a chain-coding paradigm, and faster encoding convergence (see Table 3.2 in our response to Reviewer [BiKZ]). These advances offer independent value for both INR and overfitted codec design. We welcome further suggestions to strengthen the soundness and impact of this work.

Table 1.1 Detailed RD datapoints of MoRIC on CLIC2020 (professional validation dataset).

Table 1.2 Comparison of different schemes for standard full image compression on CLIC2020.

• W3. (b) Decoding speed: MoRIC achieves the fastest CPU decoding speed among all methods in Table 4 (object compression) of the manuscript. Specifically, most of the computation in overfitted codecs lies in the entropy model and synthesis module. MoRIC addresses this by two key design choices:

  • Explicit contour representation: MoRIC skips the expensive auto-regressive decoding for latents outside target regions.

  • Progressive Concatenated Modulation (PCM): This design enables efficient region-wise reconstruction under lightweight architecture. As a result, the synthesis module requires just 262 MACs/pixel, which is approximately 81% of COOL-CHICv4 and 62% of C3, while still delivering superior RD performance.

    To improve clarity, we will incorporate the paragraphs above, the coding cost analysis (Table 2.1 (in our response to Reviewer [3Cij])), and the CLIC2020 experiment results (Table 1.1, 1.2, and 1.4) into the revised manuscript.

• W4. We would also like to respectfully clarify a possible misunderstanding: we have included both VTM-19.1 (the reference implementation of H.266) and MLIC++ (current state-of-the-art neural codec) for comparison in our manuscript. And as shown in our experiments, MoRIC outperforms VTM-19.1 (H.266) on standard image compression benchmarks such as Kodak and CLIC2020. Additionally, in single-object compression, MoRIC surpasses MLIC++ while achieving 100×–1000× lower decoding complexity. This advantage is particularly significant in high-bpp regimes.

  To address the reviewer’s concern, we will add more baselines (Table 1.3 and 1.4) and revise the manuscript to more clearly highlight the baseline choices, as follows:

Table 1.3 More baselines on Kodak dataset.

Table 1.4 More baselines on CLIC2020 dataset.

“We evaluate MoRIC on the DAVIS, Kodak, and CLIC2020 datasets, comparing it with classical codecs (VTM, the reference implementation of H.266), autoencoder-based neural codecs (EVC for low complexity; MLIC++ for state-of-the-art RD performance), and recent overfitted codecs (C3 and COOL-CHIC v4). Additional baselines are provided in the appendix.”

“MoRIC surpasses state-of-the-art neural codecs in single-object compression, and outperforms VTM on standard compression benchmarks. These results highlight progress on key challenges in overfitted codecs, including achieving high RD performance with low decoding complexity and faster encoding through region-wise coding.”

Answers to questions:

• Q1. In our response to W2, we provided a detailed breakdown of both the theoretical and practical coding costs (presented as Table 2.1 in our response to Reviewer [3Cij]).

• Q2. In our response to W4, we incorporated additional baselines (shown in Tables 1.3 and 1.4).

• Q3. We added experiments on the CLIC2020 dataset (a standard benchmark with 41 high-resolution images) and included additional MS-SSIM results on Kodak, demonstrating promising performance gains (see Tables 1.1, 1.2, 1.4, and 4.2).

Dear reviewer,

Thank you for the review.

As the discussion phase is ending soon, can you please go over the authors' response and acknowledge it?

Best,
AC