- 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
- 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.
- 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.
- 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.
- 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.
- 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
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.
| Task | w/o Check | Ours |
|---|---|---|
| Object addition | 8.85 | 9.18 |
| Appearance editing | 8.54 | 8.99 |
| Pose/view editing | 7.86 | 8.57 |
| Object replacement | 8.44 | 8.84 |
| Conditional scene generation | 8.17 | 9.27 |
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 set | Method | IA ↑ | SCA ↑ | PVA ↑ | IAF ↑ | NCP ↑ | IEQ ↑ | DCP ↑ | Overall ↑ |
|---|---|---|---|---|---|---|---|---|---|
| Shared insertion (55) | DreamOmni2 | 4.75 | 4.75 | 5.80 | 6.02 | 7.44 | 6.27 | 4.89 | 5.82 |
| Shared insertion (55) | DIRECT | 7.73 | 8.22 | 8.51 | 8.33 | 9.36 | 6.44 | 6.25 | 8.09 |
| Shared insertion (55) | DecoEdit | 9.20 | 8.98 | 9.15 | 8.56 | 9.64 | 9.13 | 8.93 | 9.10 |
| Full DecoBench (256) | DreamOmni2 | 4.72 | 6.07 | 5.87 | 5.91 | 7.63 | 7.26 | 5.37 | 6.25 |
| Full DecoBench (256) | DIRECT | 5.88 | 6.61 | 7.08 | 7.60 | 7.90 | 5.18 | 6.52 | 6.25 |
| Full DecoBench (256) | DecoEdit | 8.96 | 9.05 | 9.05 | 8.66 | 9.37 | 8.89 | 8.88 | 8.96 |
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.
Response to Reviewer cwBa (Part 3)
New Benchmark Table: Overall Results
| Model | 1D (100 cases) ↑ | 2D (100 cases) ↑ | Pose/view (64 cases) ↑ |
|---|---|---|---|
| HiDream-O1-Image-Dev | 0.9191 | 1.3336 | 0.7894 |
| HiDream-O1-Image | 1.4998 | 3.1149 | 2.2586 |
| OmniGen-v1 | 2.3735 | 2.3451 | 0.7746 |
| BAGEL | 4.2500 | 3.7280 | 2.2626 |
| OmniGen2 | 4.9603 | 4.5455 | 4.7254 |
| FireRed-Edit-1.1 | 6.4308 | 5.7937 | 4.7437 |
| FireRed-Edit-1.0 | 6.7985 | 5.4321 | 3.3713 |
| DreamOmni2 | 6.8073 | 5.9692 | 5.9896 |
| FLUX.2-klein-4B | 7.0879 | 6.2227 | 6.1365 |
| FLUX.2-dev | 7.1190 | 6.1561 | 6.9241 |
| SenseNova-U1.5 | 7.1202 | 6.9211 | 6.6451 |
| Qwen-Image-2.1 | 7.3019 | 6.0394 | 5.5811 |
| Qwen-image-edit-2511 | 7.6868 | 5.7289 | 5.9866 |
| FLUX-2-klein-9B | 7.8063 | 6.9762 | 6.2083 |
| Wan2.7-image | 8.1232 | 7.8920 | 7.4475 |
| GPT-Image-1.5 | 8.1306 | 7.1814 | 8.1808 |
| Gemini-2.5-flash-image | 8.2367 | 6.9138 | 6.1853 |
| FLUX-2-max | 8.4045 | 7.9930 | 8.2132 |
| Gemini-3.1-flash-image | 8.5649 | 7.9565 | 7.4907 |
| HY-image-v3.0 | 8.6698 | 6.9786 | 6.0026 |
| Seedream 5.0 | 8.6855 | 8.0598 | 7.4996 |
| DecoEdit (Ours) | 8.7178 | 8.0660 | 8.4237 |
| GPT-Image-2 | 8.8439 | 8.2476 | 8.3225 |
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.
| Granularity | Case / configuration | Original instruction | What text/visual input specifies |
|---|---|---|---|
| Minimal, no angle | TR001, T01 | “Place a large black leather sofa in the center foreground of reference1.” | Object and location; no target view. |
| Minimal, visual view | TR009, T05 | “Place reference2 on the circular turntable near the center of the showroom.” | Reference2 supplies identity and viewpoint. |
| Visual object/pose/box | TE001, E03 | “Add reference2 to the area marked in reference3.” | Reference2 supplies object/view; reference3 supplies location. |
| Brief visual-angle request | TE018, E09 | “In reference1, adjust the motorcycle to the angle of reference2, keeping other parts unchanged.” | Text names the operation; reference2 supplies viewpoint. |
| Coarse direction | TR059, 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 viewpoint | TR002, 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-control | TR018, 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-control | TR017, 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 contact | TE009, 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. |
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.
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.
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.
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.