acceptodds
Under review as a conference paper at ICLR 2027

DecoEdit: Beyond Flat Control for 3D-Aware Unified Image Editing

Abstract

While recent advances in multimodal image editing have achieved impressive results, current frameworks remain largely confined to “flat" 2D controls (e.g., text prompts, bounding boxes, or masks). Precise 3D-aware geometric manipulation—such as coordinating an object's location, physical scale, and 3D orientation while strictly preserving its fine-grained identity—remains an elusive challenge, typically suffering from severe detail degradation or requiring expensive per-instance fine-tuning. In this paper, we present DecoEdit, an all-in-one image editing framework that goes beyond flat control by seamlessly integrating 1D textual, 2D spatial, and 3D volumetric inputs under a unified architecture. First, to power this unified model, we establish a data curation pipeline and curate a large-scale dataset, namely DecoDataset, containing 1.6 million samples, encompassing diverse multi-dimensional (1D/2D/3D) tasks and their complex compositions. Furthermore, we establish a comprehensive image editing benchmark named DecoBench to systematically evaluate multi-task editing performance. Finally, to address the fundamental trade-off between complex spatial transformations and identity preservation, we propose a novel decoupled pipeline: DecoEdit factorizes the editing process by utilizing structural guidance from SAM-3D to anchor the target geometry, scale, and spatial orientation, while concurrently leveraging a detail-preservation path to lock in high-fidelity textures. Extensive experiments demonstrate that DecoEdit achieves state-of-the-art results. Notably, on the challenging 3D spatial pose and orientation control tasks, DecoEdit significantly outperforms leading closed-source models such as Seedream 5.0 in both geometric accuracy and identity consistency.

Primary area:
datasets and benchmarks
Keywords:
Image Editing, Unified Control, 3D Control, Unified Dataset and Benchmark

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_cwBaOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies unified image editing that simultaneously satisfies textual instructions, spatial positioning, and target poses while preserving instance appearance. The authors introduce DecoDataset, comprising approximately 1.63 million training samples across 42 control configurations, and DecoBench, containing 256 standard cases and 5 special cases. The proposed method adopts a dual-stream Mixture-of-Transformers (MoT) architecture, using the original object appearance and SAM-3D-rendered target geometry as separate conditioning inputs to mitigate the conflict between pose control and detail preservation.

Critical strengths

1.The problem is practically relevant. Simultaneously satisfying geometric constraints, preserving instance-specific details, and maintaining background consistency during object rotation, insertion, and replacement remains challenging for existing image editing systems.

2.The dataset construction pipeline is relatively comprehensive. It includes combinations of control conditions, multi-stage filtering, and instance-level identity separation between training and testing. The resulting multi-control dataset could be a valuable resource for the community.

3.The experiments provide evidence of effectiveness. In Table 4, the average pose-editing score improves from 5.50 without decoupling to 8.57 with decoupling, while PVA increases from 3.44 to 8.44. All four human annotators also prefer the proposed method for pose-editing tasks.

Critical weaknesses
  1. The novelty of the core method is not sufficiently established.

The MoT architecture has clear precedents, and the principle of using the original image for appearance conditioning and a 3D proxy for geometric conditioning has also been explored in closely related work. In particular, DIRECT (ICML 2026) already injects reference appearance, 3D proxy geometry, and background conditions through separate pathways to balance pose control and instance preservation. This directly relevant work is neither discussed nor compared against.

The paper may still offer meaningful contributions through task unification and dataset construction, but it should clarify whether the main novelty lies in the conditioning mechanism, training data, or task coverage. Comparisons against general-purpose editing models alone are insufficient to establish improvements over specialized 3D editing methods. For methods with different task coverage, comparisons could be conducted on their commonly supported tasks.

  1. The benchmark has limited coverage and insufficient evidence of practical representativeness.

Although training covers ten task categories with 1D/2D/3D controls, the standard evaluation includes only five types of 3D-configured tasks. The paper does not systematically evaluate ordinary 1D/2D editing capabilities or whether joint training degrades these abilities. Moreover, only 16 cases directly evaluate pose/view editing, which provides limited evidence for the central pose-control capability.

In addition, some qualitative examples use highly templated instructions with noticeable information redundancy. For instance, after providing a target-pose reference, the instructions may repeat requirements about angles, orientation, or pose, while adding further appearance, positioning, and scene-level constraints. Such combinations increase task complexity, but the paper does not sufficiently explain their origins in real-world editing needs or whether the constraints are necessary, unambiguous, and mutually compatible. Consequently, the benchmark's representativeness of practical editing scenarios remains insufficiently justified.

  1. The evaluation evidence does not fully support the claimed degree of superiority.

Based on the presented qualitative examples, the advantages of the proposed method over the strongest competing models are not consistently apparent, making the substantial superiority suggested by the quantitative results less convincing. Although the paper emphasizes precise geometric control and high-frequency detail preservation, its evaluation relies primarily on subjective VLM-based scores, without direct measurements of angular accuracy, spatial positioning, or local detail preservation.

The improvements over the strongest baselines are also relatively small in some cases. The Macro score is 8.97 versus 8.84 for FLUX.2 [max]. For pose editing, the proposed method scores slightly below GPT-Image-1.5 (8.57 vs. 8.61), while its PVA advantage is marginal (8.44 vs. 8.38). These differences warrant paired confidence intervals or statistical significance tests.

The human evaluation is a useful complement, but it does not include these two closest competitors. Furthermore, statistical significance in the correlation between human and VLM rankings does not establish the statistical significance of the proposed method's improvements over competing methods. The use of the same judge model for training-data filtering and final evaluation further motivates independent evaluation.

  1. The ablation studies do not adequately isolate the sources of improvement.

The decoupling ablation demonstrates substantial gains, but the paper does not clearly explain which inputs, training samples, or conditioning information are modified in the non-decoupled variant. It is therefore necessary to distinguish improvements arising from additional appearance information from those attributable to the specific decoupling mechanism.

Similarly, the joint-training ablation changes both the dataset size and task composition, making it difficult to conclude that the improvements result from cross-dimensional synergy or that joint training outperforms task-specific models. Furthermore, the appendix explicitly notes that some configurations in Table 8 are evaluated on different subsets, meaning their score differences cannot be interpreted as paired ablation gains.

  1. Inconsistencies in experimental reporting and missing implementation details require clarification.

Figures 3, 9, and 10 label the competing model as GPT-Image-2, whereas Tables 2 and 8 report GPT-Image-1.5. The cost table in the appendix additionally lists both versions. The authors should clarify which model was used in each experiment and explain why GPT-Image-2 is absent from the main quantitative comparisons.

Additional details are also needed regarding the exact initialization checkpoints, initialization of the generation branch, training objectives, training budgets for both stages, and inference configurations to support reproducibility.

Questions for authors

Compared with DIRECT and other methods that separate geometric and appearance conditioning, what is the specific methodological novelty of this work? Can the authors provide direct comparisons on commonly supported tasks?

How were the benchmark instructions constructed, and how was their practical representativeness validated? Were redundancy, ambiguity, and potentially conflicting constraints systematically examined?

Can the authors provide independent, paired evaluations against the strongest competing models, while clearly identifying the exact model versions used?

How exactly is the non-decoupled baseline constructed? Can the contribution of the decoupling mechanism be isolated under matched training data, computational budgets, and conditioning information?

11 comments
AuthorsOpenReview ↗

Response to Reviewer cwBa (Part 1)

Thank you for recognizing the practical importance of the problem, the effort involved in dataset construction, and the pose-editing results.

Q1: What distinguishes the contributions from prior MoT and decoupling work, especially DIRECT, and how do they compare on shared tasks?

While it is an algorithmic method, our work focuses on dataset construction and benchmark evaluation [Primary Area: datasets and benchmarks] — the main thrust of this paper. We differ in task coverage, model architecture, data filtering and core contributions. ICLR 2027 classifies peer-reviewed papers published on or after July 17, 2026 as contemporaneous [1], and the PMLR volume containing DIRECT was published on September 29, 2026 [2]. We have added a citation to this contemporaneous work in Related Work and compared both methods on 55 shared insertion cases, alongside the broader DecoBench evaluation [3, 4]. Both works focus on 3D-aware editing, underscoring the importance of this problem.

Our contributions are multi-dimensional training data, a unified evaluation suite, and an editor that uses complementary appearance and pose guidance across editing tasks. DecoDataset and DecoBench support joint 1D/2D/3D control, and DecoEdit evaluates their value within one model.

a. Contributions and task scope.

  • DecoDataset: 1,627,974 quality-filtered samples across 10 task families and 42 control subtasks, including 650,466 3D-configured samples. It combines 1D instructions, 2D spatial/appearance conditions, and 3D pose references, covering individual controls and their combinations.
  • DecoBench: a unified benchmark for semantic adherence, spatial placement, and 3D pose/view alignment across editing tasks and control combinations.
  • DecoEdit: a unified editor with a dual-stream MoT architecture. The original object reference preserves identity and fine details; a SAM-3D-derived proxy guides target orientation, geometry, and scale.

DIRECT is a closely related contribution that studies pose-controllable object insertion and provides a supporting data pipeline [3, 4]. We view the two efforts as complementary: DIRECT focuses on this specialized insertion setting, whereas DecoDataset and DecoBench cover a broader range of editing tasks and combinations of control conditions.

Response Table 1. Task coverage and training statistics of DIRECT and DecoDataset.

Task familyDIRECTOursControlsSamplesSubtasks
Object addition / insertionYesYes1D / 2D / 3D762,8339
Object removalNoYes1D / 2D219,4722
Object replacementNoYes1D / 2D / 3D413,7998
Attribute editingNoYes1D / 2D / 3D89,9484
ResizingNoYes1D / 2D2,4233
RelocationNoYes1D / 2D7,9773
In-place pose/view editingNoYes1D / 2D / 3D20,2684
Background editingNoYes1D / 2D75,0502
Composite editingNoYes1D / 2D14,2914
Object-conditioned scene generationNoYes2D / 3D21,9133
Total1 family10 families1D / 2D / 3D1,627,97442

b. Differences in conditioning and data construction. Even for object insertion, DecoDataset provides nine configurations (T01–T09, Table 7), covering text-only input, appearance references, bounding-box guidance, joint appearance/pose references, separate appearance and pose references, and combinations of spatial controls. These configurations allow the model to control what is inserted, where it is placed, and which viewpoint and appearance should be preserved. DIRECT uses condition-specific LoRA pathways. By contrast, DecoEdit processes text through the semantic stream and spatial, appearance, and pose information through shared visual projections and multimodal attention (Sec. 3.3.1). Our methodological contribution is to apply this unified conditioning design and role-specific supervision across a broader range of editing tasks.

DIRECT constructs insertion pairs from SA-1B and MVImgNet. Our pipeline instead follows task inversion → object/bbox/pose recovery → control combinations → role-specific instruction rewriting → final verification (Sec. 3.1.1; Fig. 1). This process allows a single edit pair to generate multiple supervision configurations. The F1–F5 cascade verifies task eligibility, edit success, target-label consistency, augmented views, and the final correspondence among the instruction, references, and target. A sample is retained only if all nine F5 criteria receive 5/5. The five-task ablation in Response Table 2 further quantifies the contribution of this final screening stage: applying F5 improves the reported scores across all five task families.

AuthorsOpenReview ↗

Response to Reviewer cwBa (Part 2)

For examples of the dataset and the construction workflow, see the DecoDataset/DecoBench preview: https://docs.google.com/presentation/d/1knm7txYlusoGlwZfC_qxA65taNKZHVO0/edit?usp=sharing&rtpof=true&sd=true (dataset preview), and the pipeline demonstration vedio: https://drive.google.com/file/d/18kv5hsbLThgzABKk4KxMa3FFplSaM5eh/view?usp=sharing (data-construction and editing pipeline video).

Response Table 2. F5 filtering ablation across five task families.

Taskw/o CheckOurs
Object addition8.859.18
Appearance editing8.548.99
Pose/view editing7.868.57
Object replacement8.448.84
Conditional scene generation8.179.27

F5 filtering ablation, Table 4, p. 9.

c. Direct comparison with DIRECT. We evaluated DIRECT on 55 overlapping object-insertion cases and on the full 256-case standard DecoBench. On the shared insertion cases, DIRECT achieves 8.09, compared with 5.82 for DreamOmni2, while DecoEdit achieves 9.10 and a higher PVA (9.15 vs. 8.51). On the full benchmark, DIRECT and DreamOmni2 both obtain 6.25, compared with 8.96 for DecoEdit.

Appendix qualitative comparison (Appendix B.9, p. 60, Fig. 17): We provide side-by-side visual comparisons of DecoEdit and DIRECT on six shared object-insertion cases—bed, bird, polar bear, plate, bottle, and motorcycle—under matched source images and reference/spatial conditions, focusing on pose alignment, spatial placement, and appearance preservation. The first two cases use separate appearance and pose references; the remaining cases use joint appearance–pose references.

Response Table 3. Direct comparison with DIRECT on shared insertion cases and the full DecoBench.

Evaluation setMethodIA ↑SCA ↑PVA ↑IAF ↑NCP ↑IEQ ↑DCP ↑Overall ↑
Shared insertion (55)DreamOmni24.754.755.806.027.446.274.895.82
Shared insertion (55)DIRECT7.738.228.518.339.366.446.258.09
Shared insertion (55)DecoEdit9.208.989.158.569.649.138.939.10
Full DecoBench (256)DreamOmni24.726.075.875.917.637.265.376.25
Full DecoBench (256)DIRECT5.886.617.087.607.905.186.526.25
Full DecoBench (256)DecoEdit8.969.059.058.669.378.898.888.96

Q2: Does DecoBench cover practical 1D/2D and pose/view editing with clear, compatible instructions?

Existing benchmarks and methods, including GIER, MagicBrush, AnyEdit, EditVal, GEdit-Bench, and EditInspector, cover a broad range of conventional text-guided and region-based image-editing tasks [8–10, 12–14]. However, systematic evaluation of explicit 3D pose/view control remains comparatively limited, particularly when pose guidance is combined with appearance references and spatial constraints. DecoBench addresses this gap by examining how visual pose guidance, appearance references, and spatial controls work together. Its standard benchmark covers five 3D-configured task families, while the training set spans 10 task families and 42 subtasks across 1D/2D/3D. We further assess the unified editor through held-out 1D/2D evaluations and an expanded 64-case pose/view evaluation.

a. Evaluation coverage. We added 100 held-out 1D cases and 100 held-out 2D cases, and added 48 pose/view cases to expand that evaluation from 16 to 64 cases. The new 1D/2D sets are disjoint from the training data. Compared with initialization, the 1D Overall score increases from 8.6449 to 8.7178, and the 2D Overall score increases from 6.7764 to 8.0660. All six applicable 1D metrics improve. Five 2D metrics improve, while IEQ remains unchanged at the reported precision. On the 64-case pose/view set, DecoEdit reaches 8.4237 Overall, compared with 8.3225 for GPT-Image-2, 8.2132 for FLUX.2 Max, and 8.1808 for GPT-Image-1.5. Their PVA scores are 8.0625, 7.8125, 7.6562, and 7.3594, respectively.

AuthorsOpenReview ↗

Response to Reviewer cwBa (Part 3)

New Benchmark Table: Overall Results

Model1D (100 cases) ↑2D (100 cases) ↑Pose/view (64 cases) ↑
HiDream-O1-Image-Dev0.91911.33360.7894
HiDream-O1-Image1.49983.11492.2586
OmniGen-v12.37352.34510.7746
BAGEL4.25003.72802.2626
OmniGen24.96034.54554.7254
FireRed-Edit-1.16.43085.79374.7437
FireRed-Edit-1.06.79855.43213.3713
DreamOmni26.80735.96925.9896
FLUX.2-klein-4B7.08796.22276.1365
FLUX.2-dev7.11906.15616.9241
SenseNova-U1.57.12026.92116.6451
Qwen-Image-2.17.30196.03945.5811
Qwen-image-edit-25117.68685.72895.9866
FLUX-2-klein-9B7.80636.97626.2083
Wan2.7-image8.12327.89207.4475
GPT-Image-1.58.13067.18148.1808
Gemini-2.5-flash-image8.23676.91386.1853
FLUX-2-max8.40457.99308.2132
Gemini-3.1-flash-image8.56497.95657.4907
HY-image-v3.08.66986.97866.0026
Seedream 5.08.68558.05987.4996
DecoEdit (Ours)8.71788.06608.4237
GPT-Image-28.84398.24768.3225

Compared with initialization, all six applicable 1D metrics improve. Five 2D metrics improve, while IEQ remains unchanged at the reported precision. On the expanded pose/view set, DecoEdit achieves an Overall score of 8.4237, exceeding GPT-Image-2, FLUX.2 Max, and GPT-Image-1.5.

b. Joint-training interference. The joint-training ablation (Figure 6, p. 9; Table 9, p. 24) is evaluated on the same standard benchmark. The complete results are provided in Response Table 12 under Q4(c).

Joint training achieves the highest Macro (8.97) and the highest scores on all seven diagnostic metrics. Removing 1D or 2D supervision reduces Macro to 7.94 or 8.38, respectively. These ablations support complementary supervision within the unified editor; the held-out 1D/2D results further document its ordinary editing capabilities.

c. Instruction construction and representativeness. Gemini-3.7-Flash rewrites the instructions after the control conditions and reference roles have been established (Sec. 3.1.1, p. 4). The prompt adapts instructions to five input configurations: text-only (v1), cropped appearance reference (v2), cropped reference with spatial guide (v3), original appearance reference (v4), and original reference with guide (v5). It requires text to describe the object and placement when those are not provided visually, describe transformations when needed, use reference3 for placement/scale/layout when useful, and not mention unavailable inputs.

Response Table 4. DecoDataset instruction examples at different levels of textual specificity.

GranularityCase / configurationOriginal instructionWhat text/visual input specifies
Minimal, no angleTR001, T01“Place a large black leather sofa in the center foreground of reference1.”Object and location; no target view.
Minimal, visual viewTR009, T05“Place reference2 on the circular turntable near the center of the showroom.”Reference2 supplies identity and viewpoint.
Visual object/pose/boxTE001, E03“Add reference2 to the area marked in reference3.”Reference2 supplies object/view; reference3 supplies location.
Brief visual-angle requestTE018, E09“In reference1, adjust the motorcycle to the angle of reference2, keeping other parts unchanged.”Text names the operation; reference2 supplies viewpoint.
Coarse directionTR059, T30“Rotate the blue and white bus in the center of the scene slightly counter-clockwise so that it appears to be turning toward the left.”Direction without numerical angle.
Explicit text-only viewpointTR002, T01“Add a large Singapore Airlines Airbus A380 aircraft to the center-left of the tarmac in a horizontal, right-facing side profile view with its nose facing toward the terminal gates on the right.”Text specifies view and placement.
Broad multi-controlTR018, T09“Add reference2 to the area marked in reference4, adjusting its angle, pose, and lighting to integrate naturally into the scene.”General pose, placement, and lighting.
Detailed multi-controlTR017, T09“Adjust its angle to a top-down view, change the pose so the shoes are parallel and pointing upwards, and modify the lighting to cast realistic soft shadows.”Explicit view, arrangement, and shadows.
Physical contactTE009, E08“Adjust the angle of the handlebars to fit the figure's hands and position the pedal to meet her foot.”Contact constraints between object and scene.
Reviewer_UN7YOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

DecoEdit unifies textual instructions, 2D spatial controls, and 3D pose references in a dual-stream Mixture-of-Transformers image editor. The authors construct DecoDataset with approximately 1.63 million samples and 42 control subtasks through condition expansion, instruction rewriting, and five-stage filtering. Appearance and geometry are conditioned separately using an original object reference and a SAM-3D-derived target-pose rendering. Evaluation on DecoBench combines multimodal-model scores and human rankings, showing particularly strong performance on pose-controlled editing.

Critical strengths
  1. Separating target orientation from source appearance addresses a meaningful conflict between geometric manipulation and instance-detail preservation.

  2. Role-specific conditions, combinatorial control generation, and staged filtering form a substantial resource, with detailed task and sample statistics.

  3. Various task-level results transparently show that its superiority is not uniform across tasks.

Critical weaknesses
  1. Pose and spatial accuracy are primarily assessed by Gemini, which also filters the training data. This may favor the same model’s quality criteria. Human evaluation helps, but covers only 50 cases and excludes strong competitors such as FLUX.2 Max and GPT-Image-1.5.

  2. Evaluation coverage is narrower than the unified-editing claim. Although training covers 42 subtasks, the standard benchmark evaluates five 3D-related task families, with only 16 pose/view-editing cases. Ordinary text editing, 2D manipulation, and multi-operation compositions are not systematically evaluated. Potential interference between jointly trained capabilities also remains unclear.

  3. The decoupling mechanism is not fully separated from its informational advantage. Decoupling is primarily implemented by providing separate appearance and pose images to shared attention. Despite the useful retrained ablation, the retained conditioning information in the non-decoupled variant is insufficiently specified.

  4. Some qualitative figures identify GPT-Image-2, whereas the main table reports GPT-Image-1.5. Model versions and corresponding generation settings should be reconciled.

Questions for authors

The paper’s main contribution appears to be a substantial data collection and curation effort, while the methodological insight is relatively limited. The training approach largely follows established practices for adapting pretrained diffusion models, including those widely used with FLUX.1-dev. The dataset can nevertheless be a valuable contribution. However, the paper does not clearly commit to releasing even a meaningful subset of the curated data. Without such a release, the community would have limited ability to reproduce the results, examine the data-construction choices, or build upon the resource. Given the incremental methodology, the strength of the paper’s contribution therefore depends substantially on a concrete data-release plan, including its scope and licensing.

8 comments
AuthorsOpenReview ↗

Response to Reviewer UN7Y (Part 1)

Thank you for recognizing the value of our role-specific conditioning and data curation. In response to the comments, we have expanded the human study and evaluation coverage, added ablation analyses, and clarified our release plan.

Q1: Could using Gemini for both filtering and evaluation bias the results, and how has the human study been strengthened?

a. We expanded the model-blind human evaluation from 50 to 100 cases and added the previously missing strong competitors. Evaluating all 23 outputs from the 15 open-source baselines, six closed-source baselines, and two DecoEdit checkpoints would require 1,150 images across the original 50 cases. Presenting so many outputs in each group would also make it harder for raters to distinguish models with similar performance and could increase annotation errors. We therefore selected five representative models from different model families for the original human study, in which raters ranked five outputs per case (50 cases × 5 outputs = 250 images). The expanded study covers 100 cases, four raters, and nine reported models, including FLUX.2 Max and GPT-Image-1.5 (900 displayed outputs per rater).

Raters evaluate instruction fulfillment, pose/view alignment, spatial placement, identity and local-detail preservation, and natural scene integration, including potential pasted-on effects and reconstruction-proxy artifacts. This human assessment complements model-based evaluation, as in prior image-editing benchmarks [1–3]. For each case, the nine reported models retain their original per-case ranks and tied midranks. We first average the rankings within each of the five task families and then take the equally weighted mean across task families to obtain Task Macro.

Individual-annotator results for the same 100 cases are reported in Appendix B.5, Table 10, p. 25. Each task family receives 20% weight in Task Macro; these individual preferences are descriptive, not per-annotator significance tests. We report per-case preference ranks for nine models.

Response Table 1. Expanded model-blind human preference evaluation on 100 cases.

Macro rankModelAdditionReplacementAppearancePose/viewScene generationTask Macro ↑
9FireRed-Image-Edit-1.12.9332.6092.0002.0161.6252.237
8Qwen-Image-Edit-25112.8672.4472.0893.2506.7503.481
7Gemini 2.5 Flash Image5.2044.1384.4914.2343.3754.289
6HunyuanImage 3.03.4213.4794.2954.8596.8754.586
5Wan 2.7 Image7.0966.4557.2866.2034.1256.233
4GPT-Image-1.56.5836.7297.1796.7034.3756.314
3SenseNova-U1.5-8B-MoT6.3176.5456.5007.2507.1256.747
2FLUX.2 Max7.0627.5117.5896.0315.7506.789
1DecoEdit6.8177.1386.1528.2347.2507.118

DecoEdit achieves the highest Task Macro (7.118). It also leads on pose/view editing, scoring 8.234, compared with 6.031 for FLUX.2 Max and 6.703 for GPT-Image-1.5.

b. Human–VLM agreement. Using the original labels from Raters 1–4, we compute Spearman's ρ and Kendall's τ_b between the VLM rankings and each rater's ranking, as well as the human-consensus ranking, across the nine reported models in each case. We then average the correlations over the 100 cases.

Response Table 2. Agreement between Gemini-based scores and human rankings.

Human referenceMean Spearman ρ ↑Holm-adjusted p (ρ)Mean Kendall τ_b ↑
Rater 10.677< 0.0010.556
Rater 20.703< 0.0010.580
Rater 30.685< 0.0010.566
Rater 40.667< 0.0010.542
Human consensus0.752< 0.0010.623

For each correlation comparison, we use a two-sided permutation test with 100,000 permutations and apply Holm correction across the five Spearman comparisons (four individual raters and the human consensus) [4]. All adjusted p-values are < 0.001. The individual-rater correlations range from ρ = 0.667 to 0.703, while the correlation with the human consensus is ρ = 0.752 (τ_b = 0.623). According to Evans [5], values of ρ between 0.60 and 0.79 indicate strong correlation. These results show substantial agreement between the VLM rankings and the independent human assessments on the evaluated cases.

AuthorsOpenReview ↗

Response to Reviewer UN7Y (Part 2)

Q2: Does the evaluation cover 1D/2D editing and pose/view control, and test for joint-training interference?

Existing benchmarks and methods—including GEdit-Bench, EditVal, EditInspector, MagicBrush, and AnyEdit—cover a broad range of general image editing, instruction following, and spatially guided manipulation [1–3, 9, 10]. However, systematic evaluation of 3D pose/view control remains relatively limited, particularly when pose guidance is combined with appearance references and spatial constraints. DecoBench is designed to address this gap by evaluating the joint use of visual pose guidance, appearance references, and spatial controls. This focus is reflected in its five 3D-configured task families, while the full training set covers 10 task families and 42 subtasks across 1D/2D/3D. We further assess the breadth of the unified editor through held-out 1D/2D evaluations and an expanded pose/view evaluation, increased from 16 to 64 cases.

a. Additional held-out 1D/2D evaluations and an expanded pose/view evaluation. We added 100 held-out 1D cases and 100 held-out 2D cases, and expanded the pose/view evaluation from 16 to 64 cases. The new 1D/2D sets are disjoint from the training data and follow the same construction and filtering procedures as DecoBench.

Response Table 3. Full results on the additional evaluation sets. The second-best results are shown in italics.

Evaluation setModelIA ↑SCA ↑PVA ↑IAF ↑NCP ↑IEQ ↑DCP ↑Overall ↑
1D, 100 casesInitialization8.58678.73678.81698.30959.01008.7600—8.6449
DecoEdit (16k)8.89428.94339.18848.71039.25338.7683—8.7178
GPT-Image-1.58.26428.48678.75858.41967.83678.6367—8.1306
GPT-Image-28.97178.96509.11598.96439.05339.0383—8.8439
2D, 100 casesInitialization6.15005.80836.75626.57147.92628.0483—6.7764
DecoEdit (16k)7.77337.58678.07397.47838.07478.0483—8.0660
GPT-Image-1.56.53006.17837.53057.31507.05387.7550—7.1814
GPT-Image-28.08087.70338.50008.53808.53658.6150—8.2476
Pose/view, 64 casesDecoEdit (16k)7.90628.70318.06258.21889.17198.59388.37708.4237
GPT-Image-1.57.17198.64067.35948.32819.15628.89067.79318.1808
GPT-Image-27.34388.93757.81258.14069.50008.82817.76678.3225

Compared with initialization, all six applicable 1D metrics improve. Five 2D metrics improve, while IEQ remains unchanged at the reported precision. On the expanded pose/view set, DecoEdit achieves an Overall score of 8.4237, exceeding GPT-Image-2 (8.3225), FLUX.2 Max (8.2132), and GPT-Image-1.5 (8.1808). The corresponding PVA scores are 8.0625, 7.8125, 7.6562, and 7.3594, respectively.

(1) Comparison with commercial systems. Closed-source commercial systems do not disclose their architectures or training costs, and may involve complex pipelines rather than standalone models, making direct comparisons difficult. Previous works, including NEP [7] and EditMGT [8], mainly compare open research models. Despite using an only 7B generation branch and 16 A800 GPUs for Stage-2 adaptation (excluding Stage-1 pretraining), DecoEdit achieves an excellent score (7.118), outperforming several closed systems.

(2) Comparison with GPT-Image-2. Released on April 21, 2026 [6], GPT-Image-2 represents a contemporaneous commercial system. DecoEdit achieves comparable 1D (8.7178 vs. 8.8439) and 2D (8.0660 vs. 8.2476) editing performance, while outperforming it on pose/view editing (8.4237 vs. 8.3225).

(3) Our model's strong performance is one of several key contributions. Beyond model performance, we develop a data curation pipeline and construct DecoDataset, comprising 1.6 million samples covering diverse compositional 1D/2D/3D editing tasks. We further establish DecoBench for systematic evaluation and propose a unified framework that enables multi-dimensional editing while preserving fine-grained object identity.

b. Joint-training ablation. Figure 6 (p. 9) and Table 9 (p. 24) compare different training mixtures on the same standard benchmark. The ablation examines whether 1D, 2D, and 3D supervision provide complementary capabilities to the same unified editor.

AuthorsOpenReview ↗

Response to Reviewer UN7Y (Part 3)

Response Table 4. Joint-training ablation across 1D, 2D, and 3D supervision.

Training mixtureIA ↑SCA ↑PVA ↑IAF ↑NCP ↑IEQ ↑DCP ↑Macro ↑
Joint 1D + 2D + 3D8.989.079.058.709.398.928.908.97
w/o 1D (2D + 3D)7.708.258.127.758.678.068.207.94
w/o 2D (1D + 3D)8.328.698.748.208.848.388.578.38

Joint training achieves the highest Macro (8.97) and the highest scores on all seven diagnostic metrics. Removing 1D or 2D supervision reduces Macro to 7.94 or 8.38, respectively. The task-level metric breakdown is provided in revised-manuscript Table 9 (p. 24). Together with the held-out 1D/2D results, these findings support complementary supervision within the unified editor and document its ordinary editing capabilities under the evaluated settings.

Q3: Does the decoupling ablation isolate role-specific supervision under identical test-time conditions?

This ablation changes the training supervision, not the test-time inputs. Both variants receive exactly the same instructions and visual conditions at inference, including appearance references. The w/o Decoupling variant removes eight training categories (163,074 samples; 10.02%), reducing the training set from 1,627,974 to 1,464,900. The images, instructions, and targets of all retained samples remain unchanged. The removed categories contain the role-specific examples that teach the model how to use the two references: the reconstructed 3D reference supplies the target viewpoint and coarse geometry, while the original or separate appearance reference preserves identity, textures, markings, and fine details that may be lost during reconstruction. Without this role-specific supervision, the model may conflate the two reference roles or rely on the degraded appearance of the 3D proxy. These effects are illustrated in Fig. 4 (p. 8). For the motorcycle, the baseline retains a front-oblique view rather than adopting the target side view, while w/o Decoupling mixes structures from different viewpoints. For the cabinet, it preserves too much of the side view instead of following the near-frontal target. Fig. 3 (p. 8) further shows that the full model can change the bird's viewpoint while preserving its plumage and markings.

Response Table 5. Eight training categories removed in the w/o Decoupling variant.

Task familyRemoved categorySamples
Pose/view editingadjust_view_with_label8,414
Pose/view editingadjust_view_without_label8,359
Pose/view editingorientation_with_ref_bbox1,076
Conditional scene generationref_extract_with_ref6,874
Reference object additionreference_add_with_ref_edit57,578
Reference object additionreference_add_with_ref_edit_bbox57,240
Reference object replacementreplace_reference_with_ref_edit11,802
Reference object replacementreplace_reference_with_ref_edit_bbox11,731
Total8 categories163,074

The two variants share the same architecture, initialization, optimizer, sampling settings, and 16,000 training iterations. Within each evaluation scope, they are tested on identical cases with identical instructions and visual conditions. The full model is not given any additional appearance reference at test time.

On the 256-case full benchmark, Macro improves from 8.095 to 8.968. Pose/view improves from 5.503 to 8.565, including PVA from 3.438 to 8.438 and IAF from 5.875 to 8.375. Addition improves from 8.651 to 9.176, appearance editing from 8.864 to 8.990, and scene generation from 8.612 to 9.268; replacement is nearly unchanged (8.847 vs. 8.841).

Reviewer_HSa5OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces DecoEdit, a unified image editing framework with 3D aware capabilities, capable of simultaneously processing text instructions, 2D spatial controls, and 3D pose information within a single model. Furthermore, the authors have constructed DecoDataset—a training dataset comprising approximately 1.6 million samples that supports diverse control modalities—as well as DecoBench, a comprehensive evaluation benchmark. To achieve this, the authors developed an "MOT" framework designed to directly accept and process heterogeneous conditional tokens. Extensive experiments demonstrate the effectiveness of this approach.

Critical strengths
  1. The motivation is valid. Joint control of text, 2D location, and 3D pose, while keeping object identity, is a real editing need. This task is meaningful.

  2. The method is reasonable. A MoT framework that takes in heterogeneous conditional tokens, and uses separate references for pose and appearance, is a sensible design for this task.

  3. I think that the dataset is the larger contribution. Building DecoDataset takes substantial work, including data construction, view augmentation, multi-stage filtering, and benchmark decontamination.

Critical weaknesses
  1. This paper does not account for inaccurate segmentation. While the pipeline verifies that object names correspond to the pixel and text content—and ensures that the final edited result remains within the bounding box—it does not validate the accuracy of the automatic segmentation itself.

  2. This paper offers only limited safeguards against inaccurate 3D reconstruction. If the mask overlap degrades, the method attempts to optimize the pose but reverts to the initial pose; additionally, a final filtering step checks the object's orientation. However, if the initial SAM-3D reconstruction is fundamentally flawed, this fallback mechanism will simply preserve the erroneous pose.

I hope that the authors could provide more details about these two points.

Questions for authors

N/A

6 comments
AuthorsOpenReview ↗

Response to Reviewer HSa5 (Part 1)

Thank you for recognizing the contribution of our dataset and for raising important questions about segmentation accuracy and 3D reconstruction reliability.

Q1: How does the pipeline detect and handle inaccurate segmentation?

a. Published benchmarks support our segmentation choices. We use the official Grounded-SAM pipeline [1] (Grounding DINO Swin-T [2] + SAM ViT-H [3]) to convert F3-verified labels into masks, without adapting the segmentation model. In zero-shot COCO evaluation with ViTDet-predicted boxes, SAM achieves 46.5 mask AP, compared with 51.0 for fully supervised ViTDet-H (Table 5 [3]). This comparison supports our choice of a strong general-purpose segmentation model without task-specific fine-tuning.

For 3D reconstruction, SAM 3D [4] achieves the best results among the compared methods on all four ground-truth geometry metrics in SA-3DAO:

Response Table 1. Published SA-3DAO reconstruction comparison reproduced from SAM 3D.

Model[email protected] ↑vIoU ↑Chamfer ↓EMD ↓ULIP ↑Uni3D ↑
TRELLIS [5]0.14750.13920.09020.21310.14730.3698
Hunyuan3D-2.1 [6]0.13990.12660.11260.24320.12930.3546
Hunyuan3D-2.0 [7]0.15740.15040.08660.20490.14840.3662
Direct3D-S2 [8]0.15130.14650.09620.21600.14050.3653
TripoSG [9]0.15330.14450.08440.20570.15290.3687
Hi3DGen [10]0.16290.15310.09370.21340.14190.3594
SAM 3D [4]0.23440.23110.04000.12110.14880.3707

Reproduced from Table 2 of SAM 3D [4].

b. Our pipeline checks segmentation-derived supervision at the level used by the editor. Masks serve as intermediate annotations rather than direct editor inputs. The editor is conditioned on text, bounding boxes, and visual references, not pixel-accurate masks: text specifies the requested edit, bounding boxes delimit editable regions without requiring contour-level annotations, and visual references provide appearance and pose guidance. This design is consistent with instruction-based [11] and region-based editing approaches [12, 13].

We compute the four box boundaries directly from the x/y extrema of the mask foreground, ensuring that each box is spatially consistent with its mask by construction. F3 verifies that the label identifies the instance changed between the before and after images and matches the edit instruction (pp. 27–28). F5 then checks whether the box covers the intended instance or relevant object part. In particular, Rules 5–6 reject boxes that target a different instance or an unrelated object/background region, or that materially miss or misalign with the designated object (p. 33, lines 1757–1761). F5 also verifies the object reference and checks whether the reference-angle image depicts the same edited object with an interpretable orientation (p. 37).

Taken together, these checks verify object identity and task-relevant coverage in the labels, bounding boxes, and visual references supplied to the editor. The quality-control process is therefore aligned with the information the editor actually uses.

c. Filtering statistics demonstrate substantial quality control. A sample is retained only when Gemini-3.7-Flash assigns a score of 5/5 to each of the nine task-specific F5 criteria (Table 11, p. 26).

Response Table 2. F2+F3 and F5 screening counts and retention rates.

Screening stageInputOutputRetention
F2+F3: label generation and screening1,122,270391,17634.86%
F5: nine criteria, including object correspondence, spatial conditions, and reference validity2,164,1511,629,36475.29%

F5 excludes 534,787 candidates (24.71%). After benchmark isolation, 1,627,974 training samples remain.

The Appendix provides representative cases generated with the intermediate segmentation-specific filter disabled (Appendix B.8, p. 58, Fig. 15). The wrong-instance motorcycle mask produces an overbroad box, while the nail-polish and fish masks omit the intended objects from their references. F5 rejects these samples through spatial-compliance or reference-validity checks, illustrating how later checks can indirectly remove segmentation failures. Consistent with these examples, the filtering ablation improves Macro from 8.37 to 8.97 (Table 4, p. 9; Sec. 4.3, p. 10). Together, the qualitative examples and quantitative ablation demonstrate how the pipeline screens segmentation-derived supervision for consistency and practical editing utility.

AuthorsOpenReview ↗

Response to Reviewer HSa5 (Part 2)

The relevant rules state:

Page 27 | F3, Check 1 (lines 1450–1451): Does the label uniquely identify the object instance that actually changed from BEFORE to AFTER?

Page 28 | F3, Check 2 (line 1465): Does the label uniquely identify the target object described in the edit instruction?

Page 33 | F5, Rules 5–6 (lines 1757–1761): A box on a different instance, an unrelated object/background region, or one materially missing or misaligned with the designated object is invalid guidance and MUST substantially reduce Rules 5 and 6.

Page 37 | F5, Rule 2 (lines 1958, 1962): A provided object-reference image must be visually coherent and semantically clear, and must identify the object's identity, appearance, or required attributes.

Page 37 | F5, Rule 3 (lines 1987, 1990): The reference-angle image must show enough of the edited object to infer orientation and present interpretable angle information for the same object.

Dataset preview: https://docs.google.com/presentation/d/1knm7txYlusoGlwZfC_qxA65taNKZHVO0/edit?usp=sharing&rtpof=true&sd=true (DecoDataset examples).

Q2: How does the pipeline guard against flawed SAM 3D reconstructions and incorrect poses?

Fallback preserves mask alignment without relaxing sample validation. If optimization degrades alignment, we retain the initial pose. Every assembled sample must pass the nine-criterion F5 check, and generated novel views also undergo F4. These checks assess reference validity, reference–target viewpoint consistency, and physical and structural plausibility.

a. We evaluated seven reconstruction models on the same 100 objects randomly sampled from the original DecoDataset, yielding 700 reconstructions in total. With model identities hidden, GPT-6.1 rated each reconstruction on four dimensions—Shape/proportions, Structure, Detail, and Integrity—using a 0–5 scale. We then mapped the equally weighted mean of the four scores to a 0–10 scale:

Total = (Shape + Structure + Detail + Integrity) / 2.

Response Table 3. Seven-model reconstruction assessment on 100 objects sampled from the original DecoDataset.

ModelShape /5Structure /5Detail /5Integrity /5Total mean ± SD /10
TRELLIS3.483.252.363.116.11 ± 1.83
Direct3D-S24.134.002.903.357.18 ± 0.78
Hi3DGen4.144.012.903.387.22 ± 0.97
TripoSG4.173.983.183.717.53 ± 1.23
Hunyuan3D-2.04.284.053.024.007.67 ± 0.88
Hunyuan3D-2.14.274.083.293.927.78 ± 1.04
SAM 3D4.284.143.334.258.00 ± 0.83

SAM 3D achieves the highest mean score (8.00/10) with a relatively low standard deviation (0.83). It leads in Structure, Detail, and Integrity, and ties for the highest Shape score.

Appendix qualitative comparison (Appendix B.8, p. 58, Fig. 16): We include side-by-side reconstructions produced by SAM 3D and the other six models for a recycling bin and a hot-air balloon sampled from the original DecoDataset. Six views under the same gray rendering expose shape and structural consistency; canonical orientations may differ across models. The published geometry benchmarks and our same-object comparison together support the use of SAM 3D as the reconstruction component in our pipeline; F4 and F5 subsequently evaluate the assembled reference–target samples.

b. F5 validates the assembled reference–target pair after pose optimization or fallback. The pose scaffold is derived from the edited target (Sec. 3.1.1). F5 verifies that the reference and target depict the same edited object with an interpretable and consistent orientation, and assesses physical and structural plausibility. F4 further checks whether the object remains a coherent 3D entity across augmented views.

Page 29 | F4, Rule 3 (lines 1554–1556): The target object must rotate as a complete, coherent 3D entity, with all sub-parts maintaining their relative spatial positions and rotating synchronously.

Page 37 | F5, Rule 3 (lines 1987, 1992): The reference must show enough of the edited object to infer orientation, and its visible content must not materially contradict its description or edit instruction.

Page 41 | F5, Rule 7 (lines 2200–2201): When a dedicated reference-angle image is used, the target object's coarse viewpoint, facing direction, and tilt must agree with it.

Page 38 | F5, Rule 4 (lines 2011–2014): Melting structures, broken geometry, severe blur, duplicated parts, ghost remnants, floating fragments, corrupted textures, or implausible hallucinations must reduce the score according to their visibility and effect on training usefulness.

AuthorsOpenReview ↗

Response to Reviewer HSa5 (Part 3)

Fallback changes only which pose candidate is used; it does not alter the sample-acceptance criteria. A sample is accepted only if all nine F5 criteria receive 5/5. A sample is rejected whenever any F5 criterion receives less than 5/5, including when the initial pose is retained.

c. Appearance–geometry decoupling reduces the dependence of fine-detail preservation on reconstruction fidelity. The 3D rendering provides the target orientation and coarse geometry, while the source image or a dedicated appearance reference preserves identity, textures, patterns, and fine details that may be lost during reconstruction. DIRECT likewise discusses proxy-appearance degradation and uses complementary appearance references [14]. Without role-specific supervision, an editor may conflate pose and appearance information (Fig. 4, p. 8). In our retrained ablation, the pose/view mean increases from 5.50 to 8.57 (Table 4). Fig. 4 illustrates improved viewpoint adherence for the motorcycle and cabinet, while Fig. 13 (p. 23) shows pose control alongside preservation of the hat's fur and leather appearance. These results support the value of complementary-reference learning beyond the choice of reconstruction model.

Pipeline demonstration video: https://drive.google.com/file/d/18kv5hsbLThgzABKk4KxMa3FFplSaM5eh/view?usp=sharing (Data-construction and editing pipeline).


The revised manuscript clarifies how mask-derived conditions are used and adds the segmentation-failure and seven-model reconstruction comparisons (Appendix B.8, p. 58). The existing F3–F5 prompts are provided in Appendix B.7.

References

[1] T. Ren, S. Liu, A. Zeng, et al., “Grounded SAM: Assembling Open-World Models for Diverse Visual Tasks,” arXiv:2401.14159, 2024.

[2] S. Liu, Z. Zeng, T. Ren, et al., “Grounding DINO: Marrying DINO with Grounded Pre-training for Open-Set Object Detection,” ECCV 2024.

[3] A. Kirillov, E. Mintun, N. Ravi, et al., “Segment Anything,” ICCV 2023, pp. 4015–4026.

[4] X. Chen, F.-J. Chu, P. Gleize, et al., “SAM 3D: 3Dfy Anything in Images,” CVPR 2026, pp. 7220–7232.

[5] J. Xiang, Z. Lv, S. Xu, et al., “Structured 3D Latents for Scalable and Versatile 3D Generation,” CVPR 2025, pp. 21469–21480.

[6] Team Hunyuan3D, S. Yang, M. Yang, et al., “Hunyuan3D 2.1: From Images to High-Fidelity 3D Assets with Production-Ready PBR Material,” arXiv:2506.15442, 2025.

[7] Z. Zhao, Z. Lai, Q. Lin, et al., “Hunyuan3D 2.0: Scaling Diffusion Models for High Resolution Textured 3D Assets Generation,” arXiv:2501.12202, 2025.

[8] S. Wu, Y. Lin, F. Zhang, et al., “Direct3D-S2: Gigascale 3D Generation Made Easy with Spatial Sparse Attention,” arXiv:2505.17412, 2025.

[9] Y. Li, Z.-X. Zou, Z. Liu, et al., “TripoSG: High-Fidelity 3D Shape Synthesis Using Large-Scale Rectified Flow Models,” arXiv:2502.06608, 2025.

[10] C. Ye, Y. Wu, Z. Lu, et al., “Hi3DGen: High-Fidelity 3D Geometry Generation from Images via Normal Bridging,” arXiv:2503.22236, 2025.

[11] T. Brooks, A. Holynski, and A. A. Efros, “InstructPix2Pix: Learning To Follow Image Editing Instructions,” CVPR 2023, pp. 18392–18402.

[12] X. Shuai, H. Ding, X. Ma, et al., “A Survey of Multimodal-Guided Image Editing with Text-to-Image Diffusion Models,” arXiv:2406.14555, 2024.

[13] Y. Lin, Y.-W. Chen, Y.-H. Tsai, et al., “Text-Driven Image Editing via Learnable Regions,” CVPR 2024, pp. 7059–7068.

[14] J. Gong, Y. Wang, Y. Lan, et al., “Direct 3D-Aware Object Insertion via Decomposed Visual Proxies,” ICML 2026, PMLR 306:35958–35975.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Global Response (Part 1)

Dear Area Chairs, Senior Area Chairs, and Program Chairs,

We sincerely thank all reviewers for their careful reading and constructive feedback, and the AC, SAC, and PC for coordinating the review process. Your suggestions have helped us improve our paper.

With only a few hours remaining, we have uploaded a further revised PDF incorporating the rebuttal and follow-up updates, with changes highlighted in blue. We would be grateful if you could review these updates and take them into consideration when making the final decision.

We have uploaded the Supplementary Material as a ZIP file. For ease of review, we have also provided the materials through the following two anonymous links:

https://docs.google.com/presentation/d/1knm7txYlusoGlwZfC_qxA65taNKZHVO0/edit?usp=sharing&rtpof=true&sd=true [Dataset and Benchmark Preview]

https://drive.google.com/file/d/18kv5hsbLThgzABKk4KxMa3FFplSaM5eh/view?usp=sharing [Pipeline Demonstration Video]

1. Main Contributions

  • DecoDataset: The first image-editing dataset supporting joint textual instructions, 2D spatial constraints, and explicit 3D pose references across diverse editing tasks. It provides 1.63M curated samples across 10 task families and 42 subtasks, through a scalable construction and five-stage filtering pipeline.
  • DecoBench: A unified benchmark for jointly assessing semantic adherence, spatial placement, and 3D pose alignment, alongside identity and detail preservation.
  • DecoEdit: Our appearance–geometry decoupling mechanism uses 3D renderings to guide target pose and coarse geometry, while separate appearance references preserve object identity and fine details.
  • With only 7B parameters, DecoEdit outperforms all 15 evaluated open-source baselines and several leading closed-source models, including Seedream 5.0.

We commit to releasing the complete datasets, benchmarks, weights, construction/filtering pipelines, and training/evaluation code upon acceptance to support further research.

2. Positive Assessments

  • We thank Reviewer HSa5 for affirming that “The motivation is valid” and “The method is reasonable.” We also appreciate the recognition that “Building DecoDataset takes substantial work, including data construction, view augmentation, multi-stage filtering, and benchmark decontamination.”
  • We thank Reviewer UN7Y for recognizing that “Separating target orientation from source appearance addresses a meaningful conflict between geometric manipulation and instance-detail preservation.” The reviewer also emphasized that “Role-specific conditions, combinatorial control generation, and staged filtering form a substantial resource, with detailed task and sample statistics.”
  • We thank Reviewer cwBa for recognizing that “The problem is practically relevant” and that the dataset “could be a valuable resource for the community.” We also appreciate the assessment that “The experiments provide evidence of effectiveness,” including the pose-editing improvements and favorable human assessments.
AuthorsOpenReview ↗

Global Response (Part 2)

3. Revisions During the Rebuttal

A. Data Quality and Training Supervision

  • Segmentation accuracy [HSa5 Q1]: Clarified how segmentation failures are identified and filtered out through mask-derived bounding boxes, reference validation, and our rigorous five-stage quality-filtering pipeline.
  • Reconstruction reliability [HSa5 Q2]: Added seven-model reconstruction comparisons on 100 objects and qualitative examples, and clarified pose fallback and subsequent quality checks.
  • Decoupling and joint training [UN7Y Q2–Q3; cwBa Q4]: Clarified the removed supervision and identical inference conditions.

B. Evaluation Coverage and Comparisons

  • Judge bias and evaluation reliability [UN7Y Q1; cwBa Q3]: Expanded the blinded human study to 100 cases, four raters, and nine models, including FLUX.2 Max and GPT-Image-1.5; added human–VLM agreement analysis and paired statistical tests.
  • Coverage and practical representativeness [UN7Y Q2; cwBa Q2]: Added 100 1D and 100 2D cases, expanded pose/view evaluation to 64 cases, and clarified instruction construction, control compatibility, and composite editing.
  • Specialized comparison [cwBa Q1]: Clarified our contributions and differences from the contemporaneous work DIRECT, and added additional quantitative and qualitative comparisons on both shared insertion tasks and the full benchmark.

C. Reproducibility and Community Reuse

  • Model versions and implementation [UN7Y Q4; cwBa Q5]: Reported comprehensive experimental results for both GPT-Image-1.5 and GPT-Image-2, and clarified initialization, training objectives, the budget, and inference settings.
  • Resource release [UN7Y Q5]: Prepared all datasets, benchmarks, model weights, and code for public release, and provided an anonymous dataset subset link for preliminary access.

4. Follow-up Feedback

  • We thank Reviewer HSa5 for carefully considering our rebuttal and additional analyses, acknowledging our “detailed and comprehensive response” and stating: “I have no further questions at this time.”
  • We thank Reviewer UN7Y for reviewing the additional experiments and recognizing the improvements in human evaluation, ordinary editing coverage, model-version reporting, and resource availability. The reviewer confirmed that our release commitment “addresses my most important concern about reproducibility and reuse” and concluded: “Given this commitment and the strengthened evaluation, I am raising my rating.”

We sincerely appreciate the constructive discussion and the opportunity to improve our work. We hope these comprehensive resources will make a meaningful contribution to the research community and provide a valuable foundation for future advances in unified, controllable image editing.

Thank you again for your time and thoughtful consideration.

Best regards,
The Authors of Submission 32048

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.