acceptodds
Withdrawn from ICLR 2027

Visual Entity Oriented Diffusion Coding for ImaGene Compression

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generative image editing produces collections of closely related images, each comprising a source image and its edited variants. Such collections impose growing demands on compression and storage. We call these collections **ImaGenes**, as shared content is inherited across variants while edits alter only localized regions, just as genes are inherited while mutations create variation. Within an ImaGene, edits operate on *visual entities*, identifiable units of image content. Independent encoding redundantly represents these shared entities, while joint compression exploiting their preservation and variation remains underexplored. To this end, we propose **ImaGene Compression**, a joint compression paradigm built upon visual entities as its basic units. It represents each ImaGene through shared entities and their variations. We realize this paradigm with **EnDiC** (Visual Entity Oriented Diffusion Codec), a training-free diffusion-based compression framework. EnDiC localizes entity changes, shares diffusion reconstruction trajectories for unchanged content across variants, and encodes the diffusion direction choices that reconstruct changed entities as incremental bitstreams. Furthermore, EnDiC adapts bit allocation across spatial regions and denoising steps to improve coding efficiency. We also introduce **ImaGeneSet**, a dataset for this task, covering entity addition, removal, and modification. High-fidelity generative editing with quality screening yields clearly defined entity changes while keeping unchanged regions spatially consistent with the source. Experiments on ImaGeneSet demonstrate that EnDiC achieves better performance on several full-reference perceptual metrics at low bitrates. Our code and data will be made publicly available upon publication.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image compression, Joint compression, Diffusion models, Generative image editing, Visual entities

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_bWieOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EnDiC for jointly compressing a source image and its edited variants. It reuses diffusion trajectories for unchanged content and encodes local changes with adaptive bit allocation. Results show improved perceptual quality at low bitrates.

Critical strengths
  1. The problem is meaningful and practically relevant.
  2. Good perceptual results at low bitrates, with several recent baselines.
  3. Useful comparisons with region and residual coding.
Critical weaknesses
  1. The entity-oriented design mostly comes down to spatial masking. The novelty over existing reference-based coding is not very clear.
  2. Only 65 image clusters, all generated using one editing model and manually screened. Generalization remains unclear.
  3. The benefit of trajectory reuse is not sufficiently isolated from other components.
  4. No edited-region evaluation or computational cost analysis.
Questions for authors
  1. How much improvement comes specifically from trajectory reuse rather than masked direction coding?
  2. Does the method still work well with less spatially consistent edits?
  3. Can the authors report edited-region quality and encoding/decoding time?
2 comments
AuthorsOpenReview ↗

Authors’ response to Reviewer bWie (1/2)

Thank you for your careful review and constructive feedback. We appreciate your time and consideration and address your comments and questions below.

W1. The entity-oriented design mostly comes down to spatial masking. The novelty over existing reference-based coding is not very clear.

1. Joint coding beyond spatial masking. We study ImaGene Compression, a new setting where a source image and its edited variants share spatially aligned content and differ in localized entity changes. Existing reference-based codecs typically establish correspondence with a reference and inject its information into the target representation. EnDiC instead restructures the transmitted information and decoding process: the source trajectory is encoded once, while the edit masks and incremental direction codes guide the diffusion decoder to reuse source states and reconstruct entity changes throughout denoising. RSBA further adapts edit budgets to change-region extent and allocates bits across denoising steps according to reconstruction benefit.

2. Added comparison with reference-based coding. We add a comparison with CLC (AAAI 2025) in Appendix A.7, Figure 18. CLC compresses individual images using retrieved references. We additionally supply the decoded source as a reference candidate for each edit in CLC. EnDiC achieves competitive LPIPS and DISTS at lower bitrates on both datasets.

W2a. Only 65 image clusters, all generated using one editing model and manually screened. Generalization remains unclear.

Our dataset size and source-image selection follow common practice in image compression evaluation. We use all images from two standard compression benchmarks, Kodak (24 images) and the CLIC2020 Professional validation set (41 images), covering natural scenes, buildings, people, and everyday objects. Each source has three edited variants generated using Seedream 5.0 Pro, covering entity addition, removal, and modification. We add detailed clarification of the existing screening procedure in Appendix A.1, where edits are selected for instruction adherence, visual plausibility, and preservation of non-target content. In future work, we plan to expand ImaGeneSet with more source images and a broader range of image editing methods.

W2b & Q2. Does the method still work well with less spatially consistent edits?

We add visualizations in Appendix A.2 (Figure 12) showing subtle color, texture, and alignment differences that remain outside the detected masks. Thus, our evaluated source-edit pairs are not perfectly spatially consistent.

For additional changes detected by semantic localization or RGB-difference cues, EnDiC includes the affected regions in the edit mask and encodes them. Larger coded regions leave less shared content to reuse, reducing bitrate savings.

Large viewpoint changes and geometric transformations remain outside our current evaluation. Evaluating and extending EnDiC to such edits is an interesting direction for future work.

AuthorsOpenReview ↗

Authors’ response to Reviewer bWie (2/2)

W3 & Q1. The benefit of trajectory reuse is not sufficiently isolated from other components. How much improvement comes specifically from trajectory reuse rather than masked direction coding?

Thank you for raising this point. We add a controlled ablation in Section 4.4, Figure 8, comparing stepwise trajectory reuse with fusion only at the final latent or RGB output. All variants retain masked direction coding and share the same source and mask streams and per-step atom counts. We re-encode the edit directions for each control under this fixed allocation, yielding identical total bpp across variants. The representative low-rate results below show the additional gains from stepwise trajectory reuse under these controlled conditions.

Kodak

FusionbppPSNR ↑LPIPS-V ↓DISTS ↓
Final RGB0.017022.6020.29670.1143
Final latent0.017022.7180.29110.1135
Stepwise (EnDiC)0.017022.8170.28460.1074

CLIC512

FusionbppPSNR ↑LPIPS-V ↓DISTS ↓
Final RGB0.017922.4570.27230.1163
Final latent0.017922.6320.26660.1136
Stepwise (EnDiC)0.017922.7150.26260.1096

W4 & Q3. No edited-region evaluation or computational cost analysis. Can the authors report edited-region quality and encoding/decoding time?

Thank you for pointing this out. We have added both regional reconstruction evaluation and computational cost analysis to the paper.

Edited-region quality. We add separate PSNR and LPIPS evaluations for edited and unchanged regions in Section 4.3 (Figure 7), with metric definitions in Appendix A.1. We also add source-image reconstruction results in Appendix A.4.

The table below provides a preview. At matched total rates, EnDiC achieves higher PSNR and lower LPIPS than independent Turbo-DDCM for the source, edited regions, and unchanged regions on both datasets. Total bitrate includes the source, all three edits, masks, and required side information.

ImaGeneSet-Kodak, total bpp = 0.03377

RegionMethodPSNR ↑LPIPS ↓
SourceTurbo-DDCM22.650.3099
SourceEnDiC25.140.2019
Edited regionTurbo-DDCM21.790.2827
Edited regionEnDiC24.670.1871
Unchanged regionTurbo-DDCM21.730.3260
Unchanged regionEnDiC23.760.2332

ImaGeneSet-CLIC512, total bpp = 0.05860

RegionMethodPSNR ↑LPIPS ↓
SourceTurbo-DDCM23.550.2391
SourceEnDiC26.080.1512
Edited regionTurbo-DDCM22.170.2361
Edited regionEnDiC24.660.1667
Unchanged regionTurbo-DDCM22.650.2540
Unchanged regionEnDiC24.670.1835

Encoding and decoding time. We have added computational cost analysis and the timing protocol in Appendix A.10. All measurements are performed on an NVIDIA RTX 4090, with mask generation excluded. The following table summarizes the encoding and decoding times per four-image cluster. EnDiC's encoding time includes the optional 80-second refinement.

ImaGeneSet-Kodak

MethodEncoding (s)Decoding (s)
EnDiC (M=122M=122)107.646.43
EnDiC-Fast (M=122M=122)8.086.45
Turbo-DDCM (independent, M=105M=105)6.916.63

ImaGeneSet-CLIC512

MethodEncoding (s)Decoding (s)
EnDiC (M=122M=122)120.356.69
EnDiC-Fast (M=122M=122)8.296.65
Turbo-DDCM (independent, M=105M=105)7.066.78

EnDiC-Fast reduces encoding time by 13.32×13.32\times on Kodak and 14.51×14.51\times on CLIC512, with minor quality degradation and nearly unchanged decoding time. For reference, Turbo-DDCM independently encodes the four images at M=105M=105.

Reviewer_dCH3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates joint compression of a source image and its generatively edited variants, referred to as an "ImaGene," where most visual content is shared and modifications are localized. The authors propose EnDiC, a training-free diffusion-based codec that extends noise-codebook-based direction coding to this setting. The method encodes a shared source reconstruction trajectory and represents edited regions through incremental bitstreams. During decoding, edit-specific latent states are fused with the source trajectory at each denoising step. Additionally, a Region- and Step-Aware Bit Allocation (RSBA) strategy adjusts coding budgets according to edit-region size and denoising stages.

The authors also introduce ImaGeneSet, a dataset containing 65 image groups, each consisting of one source image and three edited variants covering entity addition, removal, and modification. Experiments compare EnDiC against conventional, learned, and generative compression methods, as well as video codecs and region/residual coding alternatives. The results demonstrate favorable performance on several perceptual quality metrics at low bitrates, with additional ablation studies evaluating the proposed bit allocation and entity localization components.

Critical strengths
  1. Relevant and well-motivated problem setting. The paper considers joint compression of a source image and its generatively edited variants, where substantial content is shared across images. This is a practically relevant scenario, and exploiting such cross-image redundancy is a reasonable direction for improving compression efficiency.

  2. Reasonably extensive experimental comparisons. The paper evaluates the proposed method against various image and video compression approaches, including additional region-based and residual coding baselines. Ablation studies on bit allocation and entity localization are also provided. These experiments offer useful evidence of the method's effectiveness in the considered setting, although some key aspects of the proposed approach remain insufficiently validated.

Critical weaknesses
  1. Limited methodological novelty and insufficient validation of the key contribution. The proposed method builds heavily upon Turbo-DDCM, inheriting its noise-codebook-based direction coding and related region-aware guidance mechanisms. The main methodological difference appears to be the reuse of source diffusion states through mask-guided fusion during denoising. However, the paper does not provide a direct ablation isolating the benefit of this stepwise trajectory reuse, such as comparing it with fusion performed only at the final latent or RGB level under the same coding configuration. Therefore, it remains unclear whether the proposed cross-image interaction is essential to the reported improvements. Moreover, despite the emphasis on visual entities, the actual coding mechanism primarily operates through spatial masks rather than explicit entity-level representations, making the claimed novelty somewhat overstated.

  2. Limited experimental validation of the proposed entity-oriented compression setting. The evaluation relies entirely on the newly constructed ImaGeneSet, which contains only 65 image groups with carefully selected, spatially consistent local edits. While this setting is suitable for the proposed method, it remains unclear how well the approach generalizes to more diverse editing scenarios. In addition, the evaluation mainly reports full-image quality metrics, without separately analyzing reconstruction quality in edited and unchanged regions. This makes it difficult to determine whether the improvements come from better reconstruction of edited entities or simply from reusing the largely unchanged source content. Further analysis across different edit-region sizes and edit types would be important to substantiate the claimed benefits of entity-oriented compression.

Questions for authors
  1. Contribution of stepwise trajectory reuse. The main methodological novelty appears to be the reuse of source diffusion states through mask-guided fusion at each denoising step. Could the authors provide a controlled comparison against fusion performed only at the final latent or RGB level, while keeping the coding budget and other components unchanged? This would help clarify whether stepwise trajectory reuse is necessary for the reported improvements, particularly given the similarities between the direction coding mechanism and Turbo-DDCM.

  2. Effectiveness of entity-oriented compression. The current evaluation mainly reports full-image quality metrics, while the proposed method explicitly focuses on edited entities. Could the authors separately report reconstruction quality within edited and unchanged regions, and analyze performance across different edit-region sizes or edit types? This would help determine whether the improvements come from better reconstruction of edited content or primarily from preserving shared regions.

  3. Dataset representativeness and evaluation protocol. ImaGeneSet contains only 65 image groups, with manually screened edits designed to preserve non-target content. Could the authors clarify the distribution of edit-region sizes and the criteria used to select the final samples? Additionally, Appendix A.2 appears to select the temporal allocation parameter using results on the evaluation subsets. Was an independent development set used for hyperparameter selection? These details are important for assessing the generalizability and reliability of the reported results.

2 comments
AuthorsOpenReview ↗

Authors’ response to Reviewer dCH3 (1/2)

Thank you for your careful review and constructive comments. We appreciate your feedback and address your concerns and questions below.

W1a. Limited methodological novelty and insufficient validation of the key contribution. The main methodological difference appears to be the reuse of source diffusion states through mask-guided fusion during denoising.

We respectfully disagree with the assessment of limited methodological novelty. We study ImaGene Compression, a new joint compression setting for a source image and its edited variants. This setting features shared, spatially aligned content and localized entity changes. To specifically address these characteristics, we design our joint coding framework EnDiC. The key difference from Turbo-DDCM is not simply adding mask guidance, but a fundamental change in the coding protocol.

1. An entity-oriented compression formulation. ImaGene Compression organizes joint coding around preserved content and semantic entity changes. Turbo-DDCM independently encodes a complete diffusion trajectory for each image. EnDiC represents an entire ImaGene as one shared source trajectory together with multiple incremental edit streams. Entity localization supplies the spatial masks used for incremental coding, trajectory fusion, and allocation.

2. Shared-trajectory joint coding. At each denoising step, shared source states provide context for reconstructing entity changes through incremental direction codes. The edit masks are transmitted and used by the decoder for stepwise cross-image trajectory fusion. This is fundamentally different from the priority mask in Turbo-DDCM, which only spatially weights reconstruction residuals during single-image encoding and is neither transmitted nor used by the decoder.

3. Region- and step-aware allocation. RSBA further allocates incremental direction-coding capacity across edit regions and denoising steps. It adapts edit-branch budgets to change-region extent and distributes capacity across steps according to reconstruction benefit.

W1b & Q1. The paper does not provide a direct ablation isolating the benefit of this stepwise trajectory reuse. Could the authors provide a controlled comparison against fusion performed only at the final latent or RGB level, while keeping the coding budget and other components unchanged?

Thank you for the suggestion. We have added a controlled comparison of stepwise, final-latent, and final-RGB fusion in Section 4.4 and Figure 8. All variants share source/mask streams and per-step atom counts, giving identical bitrates. Final-fusion variants denoise edits independently, retaining mask-weighted direction coding and re-encoding direction codes under the same allocation. Results show that stepwise reuse outperforms both final-fusion controls.

Kodak

FusionbppPSNR ↑LPIPS-V ↓DISTS ↓
Final RGB0.017022.6020.29670.1143
Final latent0.017022.7180.29110.1135
Stepwise (EnDiC)0.017022.8170.28460.1074

CLIC512

FusionbppPSNR ↑LPIPS-V ↓DISTS ↓
Final RGB0.017922.4570.27230.1163
Final latent0.017922.6320.26660.1136
Stepwise (EnDiC)0.017922.7150.26260.1096

W1c. Despite the emphasis on visual entities, the actual coding mechanism primarily operates through spatial masks rather than explicit entity-level representations, making the claimed novelty somewhat overstated.

We would like to clarify that entity-oriented describes how semantic edits are identified and organized for joint compression. Although the codec uses a spatial mask, this mask is produced through entity-level semantic reasoning and localization, rather than solely from pixel-level differences. As detailed in Section 3.3 (Figure 4), our Entity Change Localization pipeline first uses Qwen to identify the semantic editing operation and affected entities. SAM 3 then segments these entities using semantic and spatial prompts, while RGB-difference cues supplement changes not fully covered by semantic segmentation. These masks guide direction coding, cross-image trajectory fusion, and region-aware bit allocation.

The ablation in Section 4.4 (Figure 10) shows better rate-distortion performance with combined semantic and RGB-difference masks than either alone. Zero-Bitrate Edit Fusion in Appendix A.9 (Figure 21) further demonstrates reuse and composition of encoded entity changes. In future work, this mask-based representation can be further refined to capture more fine-grained entity information.

AuthorsOpenReview ↗

Authors’ response to Reviewer dCH3 (2/2)

W2a & Q3a. Limited experimental validation of the proposed entity-oriented compression setting. Could the authors clarify the distribution of edit-region sizes and the criteria used to select the final samples?

1. Dataset size and source selection follow common compression practice. Our original evaluation uses all 24 Kodak and 41 CLIC2020 Professional validation images as sources, covering natural scenes, buildings, people, and everyday objects. We plan to expand ImaGeneSet with more source images and editing methods in future work.

2. Samples are screened for edit quality. We clarify the existing selection procedure in Appendix A.1. For each source, Seedream 5.0 Pro independently generates six candidates, two each for addition, removal, and modification. We select three based on instruction adherence, visual plausibility, and preservation of non-target content, excluding severe semantic artifacts and very small edited regions.

3. We add edit-region size statistics. Appendix A.2 (Figure 11) reports mask coverage across all edited variants, using combined semantic and RGB-difference masks before downsampling. Most edited regions occupy less than 20%20\% of the image, with larger edits also represented.

W2b & Q2. The evaluation mainly reports full-image quality metrics, without separately analyzing reconstruction quality in edited and unchanged regions. Could the authors separately report reconstruction quality within edited and unchanged regions, and analyze performance across different edit-region sizes or edit types?

1. Edited and unchanged regions. We add separate PSNR and LPIPS comparisons in Section 4.3 (Figure 7), with metric definitions in Appendix A.1. At matched total rates, EnDiC's perceptual gains extend to edited regions on both ImaGeneSet-Kodak and ImaGeneSet-CLIC512. The table below provides a preview. Total bitrate includes the source, all three edits, masks, and required side information.

ImaGeneSet-Kodak, total bpp = 0.03377

RegionMethodPSNR ↑LPIPS ↓
SourceTurbo-DDCM22.650.3099
SourceEnDiC25.140.2019
Edited regionTurbo-DDCM21.790.2827
Edited regionEnDiC24.670.1871
Unchanged regionTurbo-DDCM21.730.3260
Unchanged regionEnDiC23.760.2332

ImaGeneSet-CLIC512, total bpp = 0.05860

RegionMethodPSNR ↑LPIPS ↓
SourceTurbo-DDCM23.550.2391
SourceEnDiC26.080.1512
Edited regionTurbo-DDCM22.170.2361
Edited regionEnDiC24.660.1667
Unchanged regionTurbo-DDCM22.650.2540
Unchanged regionEnDiC24.670.1835

2. Edit-region sizes and edit types. The new analysis in Appendix A.5 shows comparable rate savings across edit sizes on ImaGeneSet-Kodak relative to independent Turbo-DDCM. Addition and modification show slightly greater savings than removal.

Q3b. Was an independent development set used for hyperparameter selection?

No independent development set was used. We evaluated five candidate values of the temporal allocation exponent, p∈{0.5,0.7,1.0,1.3,1.5}p\in\lbrace 0.5,0.7,1.0,1.3,1.5\rbrace, on both ImaGeneSet-Kodak and ImaGeneSet-CLIC512, and selected p=1.3p=1.3 as a practical compromise between reconstruction fidelity and perceptual quality. The selection procedure and comparative results are reported in Appendix A.3 (formerly Appendix A.2, following the addition of a new subsection).

Reviewer_7zDDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes EnDiC, a joint codec for a source image and its localized generative edits. It shares the source diffusion reconstruction trajectory, transmits edit masks and signed noise-codebook selections for changed regions, and allocates bits across regions and denoising steps. Validation uses 65 four-image collections derived from Kodak and CLIC sources, with image/video codec comparisons, source-reusing crop/residual baselines, and allocation ablations. The reported benefits primarily concern low-rate perceptual reconstruction; DCVC-UF remains stronger on PSNR/MS-SSIM.

Critical strengths
  1. A concrete and useful joint-coding mechanism. Sharing the source trajectory outside edited regions exploits redundancy across related images while retaining separately coded modifications (pp.4–6, Eqs.4–13). This is a meaningful compression contribution for the defined setting.
  2. Credible accounting and nontrivial comparisons. The reported rate includes source/edit streams, masks, schedules, headers, and padding; the GVC1D estimated-rate exception is disclosed (Appendix A.1). Appendix A.4 adapts nine codecs to source-reusing crop/residual coding, and Figure 7 includes uniform-allocation controls. These comparisons support the reported low-rate perceptual benefit beyond simply avoiding repeated backgrounds, while DCVC-UF remains stronger on PSNR/MS-SSIM.
Critical weaknesses
  1. Entity fidelity lacks systematic region-specific quantitative evaluation. Figures 6, 13, and 14 provide qualitative edited-region examples, but the quantitative full-reference metrics average complete source and edited images, including unchanged regions, and Patch FID pools patches (p.7; Appendix A.3). On 65 manually screened, four-image collections from one editor, these aggregate results do not establish how reliably small, semantically important modifications survive compression across the dataset. This matters directly to the entity-oriented claim, although it does not invalidate the whole-image gains or the deliberately restricted task.
  2. The smaller method gains are difficult to reproduce and interpret. The 80-second refinement cap is stated, but candidate-move details, the joint codec's bitrate-dependent tail rule, and crop/residual adaptation settings are incomplete (Appendices A.2/A.4). Hardware and evaluation counts are also needed to interpret the time-limited search. Appendix A.2 appropriately disclaims statistical significance for its small single-run differences; the allocation/mask gains likewise need clearer numerical support. This limits confidence in the incremental contribution, rather than establishing that the existing controls are absent or unfair.
Questions for authors
  1. Assessment-critical analysis of existing results: At matched total rates, what are the reconstruction results separately for source images, edited regions, and unchanged regions? Do the perceptual gains persist inside the edit masks? These results could resolve my main concern about entity fidelity.
  2. Implementation clarification: What candidate moves and stopping rules are used in refinement, which denoising steps are coded at each rate, and how are crops and signed residuals represented and allocated bits in the adapted baselines? Please provide the settings underlying the reported results and any available repeated-run variation.
2 comments
AuthorsOpenReview ↗

Authors’ response to Reviewer 7zDD (1/2)

Thank you so much for taking the time to read our paper and share your thoughtful feedback. We respond to your comments below and hope these replies address your questions and concerns.

W1a & Q1. Entity fidelity lacks systematic region-specific quantitative evaluation. At matched total rates, what are the reconstruction results separately for source images, edited regions, and unchanged regions? Do the perceptual gains persist inside the edit masks?

We have added evaluations for source-image reconstruction in Appendix A.4, and for edited and unchanged regions in Section 4.3, Figure 7. The perceptual gains do persist both inside and outside the edit masks.

1. Preview of the quantitative results. The table below gives a preview of region-specific results. Bitrate includes the source, all three edits, masks, and required side information. At these rates, EnDiC achieves better result than independent Turbo-DDCM compression in every part. Regional metric definitions and mask weighting are detailed in Appendix A.1.

ImaGeneSet-Kodak, total bpp = 0.03377

RegionMethodPSNR ↑LPIPS ↓
SourceTurbo-DDCM22.650.3099
SourceEnDiC25.140.2019
Edited regionTurbo-DDCM21.790.2827
Edited regionEnDiC24.670.1871
Unchanged regionTurbo-DDCM21.730.3260
Unchanged regionEnDiC23.760.2332

ImaGeneSet-CLIC512, total bpp = 0.05860

RegionMethodPSNR ↑LPIPS ↓
SourceTurbo-DDCM23.550.2391
SourceEnDiC26.080.1512
Edited regionTurbo-DDCM22.170.2361
Edited regionEnDiC24.660.1667
Unchanged regionTurbo-DDCM22.650.2540
Unchanged regionEnDiC24.670.1835

2. New source-only analysis of temporal allocation. We also add Appendix A.4, Figure 14 to compare source-image reconstruction results with Turbo-DDCM. EnDiC achieves better result.

3. Qualitative evidence from the original submission. Figure 6 and Appendix A.8 already show better preservation of edited details, including the scarf, rope loop, and electric scooter.

W1b. Aggregate results do not establish how reliably small, semantically important modifications survive compression across the dataset.

We add an analysis by edit size in Appendix A.5. For Kodak edits covering less than 8%8\% of the image, EnDiC achieves BD-rate savings of 64.6%64.6\% on PSNR and 59.5%59.5\% on LPIPS relative to independent Turbo-DDCM. This size-group analysis measures quality over whole edited images.

The original visual comparisons also show small, semantically meaningful changes, such as the recolored face-paint ring in Figure 6 and the added rope coil in Appendix A.8, Figure 19. EnDiC preserves the new color and the rope structure at low bitrates.

W1c. The evaluation is on 65 manually screened, four-image collections from one editor.

Our dataset size and source-image selection follow common practice in image compression evaluation. We use all images from two standard compression benchmarks, Kodak (24 images) and the CLIC2020 Professional validation set (41 images), as source images. Together, they cover natural scenes, buildings, people, and everyday objects, ensuring a varied set of source images for the edits. In future work, we plan to expand ImaGeneSet with more source images and a broader range of image editing methods.

AuthorsOpenReview ↗

Authors’ response to Reviewer 7zDD (2/2)

W2a & Q2a. The smaller method gains are difficult to reproduce and interpret. Which denoising steps are coded at each rate, and how are crops and signed residuals represented and allocated bits in the adapted baselines?

1. Magnitude of the gains. Below is a preview of the original ablations in Figures 9 and 10 on ImaGeneSet-CLIC512 at 0.059 total bpp.

EnDiC achieves better result than all five variants, showing that all methods can improve reconstruction.

VariantPSNR ↑LPIPS ↓
Uniform spatial allocation24.9120.1750
Uniform temporal allocation24.8450.1778
Uniform spatial and temporal allocation24.8300.1783
Semantic-only mask24.6480.1758
RGB-only mask24.7740.1790
EnDiC (Ours)24.9310.1745

2. Interpretation. RSBA allocates coding capacity according to edit size and denoising stage. Entity Change Localization combines semantic entity coverage with RGB-change cues. All controls retain joint coding, so the gains quantify the incremental contribution of allocation and localization.

3. Reproducibility. We have expanded the implementation details in Appendix A.1.

As for the coded denoising steps, Table 1 lists the coded prefixes at all seven presets. EnDiC and Turbo-DDCM share the 30-step schedule and automatic DDIM-tail rule. RSBA changes atom counts within the coded prefix.

As for crop and residual coding in the adapted baselines, each codec encodes the source once and all three edits at the same preset. Edits use semantic-mask bounding crops or full-image residuals. Signed residuals relative to the decoded source are mapped to [0,1][0,1] by (Ii−I^0+1)/2(I_i-\widehat I_0+1)/2. Crop coordinates, padding, and residual reconstruction are specified in Appendix A.1.

W2b & Q2b. Hardware and evaluation counts are also needed to interpret the time-limited search. What candidate moves and stopping rules are used in 80s refinement?

Thank you for raising these concerns. We have expanded Appendix A.10 with search settings.

In 80s refinement, each candidate removes 16 atoms in stage one or 8 in stage two from one denoising step, then uses the freed bits to add as many atoms as possible to another. Updates are accepted only when terminal source-latent MSE decreases. The two stages have deadlines of 40 and 80 seconds, with early stopping if a completed neighborhood yields no improvement.

At M=122M=122, the search averages 58.7/65.5 candidates and 3.9/4.7 accepted updates per source image on Kodak/CLIC512. Measurements use an RTX 4090 (24 GB), four CPU threads, and FP16, with timings averaged over multiple runs.

In this revision, we add EnDiC-Fast to improve encoding efficiency. It uses a shared precomputed schedule and achieves 13.32×13.32\times/14.51×14.51\times faster joint encoding on Kodak/CLIC512 with minor quality degradation.

See Appendix A.10 for detailed comparisons.

W2c. Appendix A.2 appropriately disclaims statistical significance for its small single-run differences; the allocation/mask gains likewise need clearer numerical support.

Thank you for this comment. We already provide concrete effect sizes to clarify the allocation and mask ablations in Figures 9 and 10, which show that both spatial and temporal allocation improve the rate-distortion tradeoff, and that combining semantic and RGB-difference masks performs better than using either alone.

We additionally ran EnDiC with three different random seeds at M=122M=122, using fixed masks and the full 80-second refinement. The mean and sample standard deviation of dataset-average PSNR and LPIPS-VGG below show stable reconstruction quality across seeds.

DatasetbppPSNR (mean)PSNR (std)LPIPS (mean)LPIPS (std)
ImaGeneSet-Kodak0.01703722.8000.00710.2850.0002

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.