acceptodds
Under review as a conference paper at ICLR 2027

BRIDGE: Background Routing and Isolated Discrete Gating for Coarse-Mask Local Editing

Abstract

Coarse-mask local editing must preserve the background while allowing new content to depart from the mask contour. Existing mask-conditioned editors often entangle these goals and treat the coarse mask as a shape prior. We introduce **BRIDGE, a dual-path architecture** with a full-canvas Main Path, a local Subject Path, and a **learned binary gate** that routes subject-token positional embeddings between subject-centric and background-anchored coordinates. On Qwen-Image-Edit-2511, BRIDGE keeps the mask outside the DiT backbone and trains both targets with independent noise. On BRIDGE-Bench, it achieves **0.503 local SigLIP2-T, 0.902 DINO, and 0.175 DreamSim**, outperforming ACE++ and FLUX.1-Fill in local precision. We then instantiate BRIDGE on **FLUX.2 [klein] 9B Base** with full-transformer training and subject-reference conditioning. Dense BBox supervision preserves Sub-image quality; **optional mask-cropped inference** removes out-of-mask Sub tokens, tightening user support without training a sparse model that overfits the mask silhouette. On supported ICE-Bench tasks, FLUX.2 BRIDGE averages **0.5941 over four local edits and 0.6254 over subject-guided edits**, leading the compared open-source models on the latter. Across Qwen and FLUX.2, BRIDGE separates support localization from subject geometry and improves local generation while preserving the surrounding scene.

Primary area:
generative models
Keywords:
coarse-mask local image editing, diffusion transformers, mask-shape bias, positional-embedding routing, subject-reference conditioning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_mc1MOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses editing with coarse masks through global and local branches. It learns binary gates that switch the coordinates of local tokens between two sets of positional encodings. The method is implemented on Qwen and FLUX.2 and evaluated for local generation quality, background preservation, and performance across benchmarks.

Critical strengths

Using positional encoding routing to coordinate local generation and background blending is a design worth investigating. Comparisons using the same backbone support the benefits of gating for local quality, and some coarse-mask editing examples demonstrate natural object shapes and integration with the surrounding scene.

Critical weaknesses
  1. The paper attributes its gains to mitigating “mask shape bias,” but the default pipeline first converts the mask into a bounding box, already removing the original contour information. Table 4 supports the benefits of gating for generation quality, and Appendix G.6 demonstrates adaptability to mask variations. However, these results do not sufficiently distinguish the respective contributions of bounding-box preprocessing and gating to mitigating shape bias.
  2. The main FLUX.2 comparisons use rewritten prompts, while the baseline scores come from the original evaluation. The rewriting process also accesses the source image, mask, and reference image, potentially affecting task difficulty. Particularly when the reported margins are small, the differences cannot be directly attributed to BRIDGE.
Questions for authors

Do the main improvements persist when prompts and input processing are matched? Which specific shape errors that remain after bounding-box preprocessing does the gating mechanism correct?

Reviewer_wW1vOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper jointly denoises the global Main Path and the local Subject Path in shared DiT, and selects between the positional encodings of subject coordinates and the coordinates of the corresponding Main regions via layer-wise, token-wise binary gates. The authors implement BRIDGE on Qwen-Image-Edit-2511 and FLUX.2, respectively, and present results for BRIDGE-Bench, MagicBrush, ICE-Bench. The FLUX.2 extension adopts dense BBox training and supports cropping Sub tokens outside the mask during inference.

Critical strengths
  1. The problem has practical significance. Users usually provide rough masks. How to prevent their contours from restricting the generated subject while keeping the background is a practical issue in local image editing.

  2. The experiments are relatively comprehensive. The authors provide results in BRIDGE-Bench, MagicBrush, and ICE-Bench across two backbones, which strengthens the credibility of the method.

  3. The method design is linked to the motivation. The dual pathway and learnable PE routing provide a concrete implementation for decoupling spatial support and subject geometry.

Critical weaknesses
  1. The necessity of the discrete gate. The fixed PE ablation supports adaptive routing, but the author has not explained why binary selection outperforms soft routing. They also need to prove whether the gate can separate the foreground geometry and background fusion.

  2. The experiments on FLUX.2 have changed many conditions, such as backbone, training approach, reference image conditioning, and prompt, and it is difficult to determine which factor specifically brings the improvement.

  3. The mask-shape bias lacks quantitative verification. The text-image and perceptual similarity within the BBox cannot directly measure the extent to which the generated contour is affected by the mask. Although mask intervention is included in the appendix, the cases is too small to support general conclusions.

  4. The performance of Qwen on the public benchmarks lags noticeably behind ACE++ on VLLM-QA for object addition and text generation, and FLUX.2 only leads Emu3.5 by 0.0014 in subject-guided editing, which is small.

  5. In Figs. 8, the local DINO is 0.532 and the global DreamSim is 0.796, while those in Table 4 are 0.902 and 0.054. The differences are not explained.

  6. The novelty of the routing. PosBridge uses PE transplant, and GrounDiT uses joint denoising of full images and local patches. So the authors need to illustrate the contribution and benefits brought by learnable routing.

  7. Figure 4 and Figure 5 contain many stacked blocks and workflow elements, especially the pipeline at the bottom of Figure 4. Although the caption explains that "white-background Sub preview" is white-image latent padding, the pipeline still needs to clearly distinguish model-generated content, manually padded regions, and their relationship to the final Main output.

  8. There are multiple issues in Figure 4 and Figure 5, including misaligned text and inconsistent font types and sizes.

Questions for authors

Most of my questions are in the weaknesses, and I hope the authors also answer:

  1. Data Split: Does Grid search use an independent validation set? Are training and test sets grouped by the original source image?

  2. Provide some quantitative results on changing the mask contour, area, and BBox.

  3. Show some failure cases. Although the paper discusses some limitations, it still lacks qualitative results.

3 comments
AuthorsOpenReview ↗

EMU3.5 is not a mask edit model

Although Emu3.5 scores slightly higher on BRIDGE's five trained task types (0.6104 vs. 0.6080), it does not reliably constrain edits to the supplied mask in our evaluation.

FLUX.2 Base [2] has no mask input and received none in our test. The evaluated official Emu3.5 [10] scripts also provide no dedicated mask-input procedure. We therefore supplied the mask as an additional image alongside the source image and optional subject reference. We appended a single sentence to the original rewritten prompt: "Picture 2 is the mask" or, when a subject reference was included, "Picture 3 is the mask." We then passed this augmented prompt and the input images to Emu3.5.

Full ICE-Bench task scores ↑. We report the recorded scores for all nine tasks (1,976 examples), including subject removal and tasks absent from BRIDGE's training. The mean gives equal weight to Tasks 19, 21, 22, 29, and 30 (805 examples). LB denotes latent blending [9]; ACE++ [12] uses rewritten prompts.

TaskNEmu3.5FLUX.2 BaseBase + LBACE++BRIDGE DenseBRIDGE Sparse
17 Inpainting4010.63320.65250.64720.63890.65060.6474
18 Outpainting4730.57130.62640.61620.61060.58010.5833
19 Local Subject Addition2090.63870.63690.62670.62240.63730.6374
20 Local Subject Removal2020.61240.60900.60940.53980.59060.5914
21 Local Text Render1630.56320.55220.49970.56670.55640.5575
22 Local Text Removal1960.61530.60760.59270.58700.59220.5945
29 Subject-guided Inpainting1400.57270.59150.58210.55920.57330.5731
30 Virtual Try On970.66240.68390.68390.65440.67750.6774
31 Face Swap950.47670.64360.64040.56750.57000.5706
Trained-task mean8050.61040.61440.59700.59800.60730.6080

Emu3.5 exceeds BRIDGE Sparse by 0.00246 (0.40%) on this five-task mean. Unmasked Base also scores higher. Task scores alone therefore do not establish mask control.

Background preservation and edit localization: equal means over the same five tasks. Background L1 is the mean smoothed RGB difference outside the target mask on a common grid, normalized by 255.

ConfigurationBackground L1 ↓Edit-region IoU ↑Changes inside mask ↑Outside area / image ↓
Emu3.50.047930.333040.06%11.82%
FLUX.2 Base0.032780.319047.07%7.07%
Base + LB0.010210.404090.38%0.81%
BRIDGE Sparse0.010390.471990.07%0.35%

Despite receiving a mask, Emu3.5 has IoU close to unmasked Base (0.3330 vs. 0.3190), with more outside-mask changed area (11.82% vs. 7.07%) and higher background L1 (0.04793 vs. 0.03278). For subject addition, only 36.19% of Emu3.5's detected changes lie inside the mask, versus 91.10% for BRIDGE. A mask editor must control where content is generated and reduce edits outside the requested region. The tested Emu3.5 interface does not meet this requirement reliably.

Relative to Emu3.5, BRIDGE improves IoU by 41.7%, reduces outside-mask changed area by 97.1%, and lowers background L1 by 78.3%.

How we extract the edit masks. Target masks are ICE-Bench annotations [1]. We adapt SR-Edit's region-identification pipeline (ECCV 2026) [11]: align source and output to a common grid (longest side at most 512 pixels), compute mean absolute RGB differences, apply Gaussian smoothing, threshold at 16/255, perform morphological opening/closing, and remove small connected components. We retain all remaining components, including those outside the target mask. Our fixed threshold replaces SR-Edit's Otsu threshold. These are detected change regions, not semantic object masks. Inside fractions exclude empty detections; IoU counts them as zero.

References

[1] Pan et al. ICE-Bench: A Unified and Comprehensive Benchmark for Image Creating and Editing. ICCV, 2025.

[2] Black Forest Labs. FLUX.2 [klein] 9B Base Model Card. 2026.

[9] Avrahami, Fried, and Lischinski. Blended Latent Diffusion. ACM Transactions on Graphics, 2023.

[10] Cui et al. Emu3.5: Native Multimodal Models are World Learners. 2025.

[11] Wang, Chen, Jiang, and Zhu. SR-Edit: Region-Aware Image Editing via Self-Refinement. ECCV, 2026, pp. 226-244.

[12] Mao et al. ACE++: Instruction-Based Image Creation and Editing via Context-Aware Content Filling. ICCV Workshops, 2025.

AuthorsOpenReview ↗

Why discrete is so important

We use discrete routing to keep each token at one of two designated positional references. This design requirement would remain even if continuous routing achieved a higher benchmark score.

OmniGen2 separates image identity from spatial coordinates, assigning all tokens of an image a shared identity coordinate [7, Sec. 3.3]. Qwen-Image-Layered similarly uses an explicit layer index in Layer3D RoPE [13, Sec. 3.2]. These conventions motivate treating image/layer identity as a reference to select.

BRIDGE therefore selects either the original Subject PE or the corresponding Main PE for each routed token. A soft mixture of the two encodings does not enforce this exact selection. For a future multilayer extension, coordinate interpolation from layer ID 0 to 4 would also pass through IDs 1, 2 and 3, which may already denote other layers. A fractional layer ID has no prescribed meaning as a mixture of the chosen layer identities. Binary selection switches directly between the designated references, while attention and image synthesis remain continuous.

Trained comparison. We also completed a trained continuous-routing control. Discrete STE routing gives better point estimates on 6/9 metrics with Sparse/Mask inference and 8/9 with Dense/BBox inference.

We independently trained both variants from FLUX.2 [klein] 4B Base using the same data, BBox Subject support, objective, gate architecture and training hyperparameters. The continuous control uses its sigmoid value to mix the two RoPE encodings throughout training and evaluation; the discrete model uses binary forward selection with STE gradients.

We evaluate the final checkpoints on matched BRIDGE-Bench inputs under both inference interfaces. Local metrics use the edit bounding box; global metrics, including L1, compare the complete output with the edited target. Sig-T/I denote SigLIP2 text/image similarity, and DSim denotes DreamSim.

Gate / interfaceLocal Sig-T ↑Local DINO ↑Local DSim ↓Local Sig-I ↑
Discrete / Sparse0.50600.87790.19650.9046
Continuous / Sparse0.50060.87830.19830.9040
Discrete / Dense0.50790.87410.20250.9032
Continuous / Dense0.50040.87570.20330.9025
Gate / interfaceGlobal Sig-T ↑Global DINO ↑Global DSim ↓Global Sig-I ↑Global L1 ↓
Discrete / Sparse0.77590.98850.04840.97310.0264
Continuous / Sparse0.77730.98790.04960.97160.0262
Discrete / Dense0.77440.98790.05120.97220.0266
Continuous / Dense0.77430.98730.05210.97050.0267

These results provide additional empirical support for the design: discrete routing improves local SigLIP2-T, DreamSim and SigLIP2-I under both interfaces. Continuous routing has slightly higher local DINO in both, and higher global SigLIP2-T and lower L1 with Sparse inference.

The final comparison has already been added to Section 4.3, "Continuous versus discrete routing," of the revised manuscript. The appendix, "FLUX.2 4B gate ablation and analysis," explains the reference-identity argument and gives the training/evaluation protocol, continuous control and intermediate-checkpoint results.

References

[7] Wu et al. OmniGen2: Towards Instruction-Aligned Multimodal Generation. 2025.

[13] Yin et al. Qwen-Image-Layered: Towards Inherent Editability via Layer Decomposition. 2025.

AuthorsOpenReview ↗

Follow-up before the discussion deadline

Dear Reviewer,

With only about three days remaining before the AoE discussion deadline, could you please review our latest responses and additional experimental results and reply as soon as possible? Please let us know which concerns, if any, remain unresolved so that we can address them and continue the discussion before the deadline.

If our responses and new evidence have resolved your concerns, would you consider revising your rating upward to reflect your updated assessment?

Thank you for your time and feedback.

Reviewer_4SP6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

BRIDGE addresses coarse-mask local image editing through jointly denoised full-canvas Main and local Subject paths. A learned binary gate selects subject-centric or background-anchored positional embeddings for Subject tokens. The paper implements the approach with Qwen and extends it to FLUX.2 with additional training and reference conditioning. Validation includes a constructed benchmark, MagicBrush, and ICE-Bench. Matched Qwen single-path, fixed-position, and learned-routing controls support improvements in local semantic alignment and image similarity; further experiments examine mask interventions and the FLUX extension.

Critical strengths
  1. The main controlled comparison supports a useful improvement. With the same reported backbone, data, and training budget, local SigLIP2-T/DINO/DreamSim change from .464/.870/.239 for the single-path baseline and .463/.875/.237 for fixed-PE BridgePath to .503/.902/.175 for BRIDGE (Table 4, p.8; matching details, p.13). Improvement in both semantic alignment and image similarity is consequential evidence for the learned policy.
  2. External local-edit evaluation supports practical relevance. MagicBrush tight-crop results improve over Q-Control on all four reported crop metrics (Table 11). Together with the interpretable separation of localization support from object geometry, this gives the contribution evidence beyond the constructed benchmark.
Critical weaknesses
  1. The strongest mask-robustness interpretation exceeds the quantitative evidence. Tables 1/4 evaluate bbox-local similarity, rather than sensitivity to changes in mask geometry. Appendix G.6 includes four cases, each evaluated under four mask conditions, and Appendix G.3 includes an irregular-mask comparison. These examples do not quantify semantic success, shape stability, and outside-support changes across a defined intervention set. Thus, the learned policy's task benefit is better supported than its claimed mask-shape robustness or token-specific specialization.
  2. The provenance of the main evaluation needs reconciliation. The selected grid-search configuration reports local DINO .5317 and global DreamSim .7963 (pp.15–16), whereas Tables 1/4 report .902 and .054. Dataset, checkpoint, normalization, or evaluator differences are not explained; these values cannot yet be treated as directly comparable or contradictory. The hyperparameter-selection split and source-level train/test separation also need to be explicit. These are fixable reporting concerns, but they affect confidence in the evidence supporting acceptance.
Questions for authors
  1. Evaluation clarification: Which sample sets, checkpoints, metric implementations, and normalization conventions produce the grid-search and main-table values? Which split was used for selection, and were training and evaluation separated by original source image with near-duplicate checks?
  2. Claim clarification: Which reported measurements establish stability under changes in mask geometry, beyond the illustrated examples? If such aggregate measurements are unavailable, can the robustness and token-specialization claims be limited accordingly? Quantitative intervention evidence would strengthen the recommendation.
2 comments
AuthorsOpenReview ↗

Follow-up before the discussion deadline

Dear Reviewer,

With only about three days remaining before the AoE discussion deadline, could you please review our latest responses and additional experimental results and reply as soon as possible? Please let us know which concerns, if any, remain unresolved so that we can address them and continue the discussion before the deadline.

If our responses and new evidence have resolved your concerns, would you consider revising your rating upward to reflect your updated assessment?

Thank you for your time and feedback.

AuthorsOpenReview ↗

Cross-split near-duplicate screening

Our SSCD-based screening and full-image review found no repeated input-target editing pairs among the retrieved candidates.

Source-image IDs and URLs are disjoint across the training and test splits. We additionally compared masked regions from both images of each editing pair using SSCD [4]. We cropped each region to its tight bounding box, filled pixels outside the mask with gray, and resized it to 320 × 320. Cosine similarity of at least 0.75 retrieved 10 candidate image pairs involving four test examples.

We reviewed the full images and their associated edits. We count two editing examples as duplicates only when both their inputs and targets repeat. The robot candidate shares a bicycle scene but involves different edits; the balloon candidate has different backgrounds. The threshold selects candidates for review rather than automatically identifying duplicates.

The method and findings are now included in Appendix B, “Cross-split content screening.”

[4] Pizzi et al. A Self-Supervised Descriptor for Image Copy Detection. CVPR, 2022.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.