Text-to-image models often struggle to synthesize correct occlusion relationships among multiple objects, especially in densely overlapping regions. Many training-free layout-guided methods enforce 2D spatial constraints but do not explicitly resolve depth-dependent attention competition, which can cause concept mixing and implausible occlusion. To address this problem, we propose DepthArb, a training-free framework that formulates occlusion generation as attention arbitration within a unified denoising trajectory. DepthArb employs two core occlusion-control mechanisms: Attention Arbitration Modulation suppresses background-object attention within foreground support according to relative depth, while Spatial Compactness Control limits attention dispersion to preserve object coherence. Because interference varies during generation, Occlusion Conflict Estimation constructs a shared spatial conflict field to adaptively weight both objectives. Through a unified spatial-text attention interface, DepthArb operates on U-Net cross-attention and the image-to-text component of MMDiT joint attention without model retraining. We further introduce OcclBench, a benchmark with continuous relative-depth specifications and occlusion-specific evaluation metrics. Experiments on OcclBench and public benchmarks show that DepthArb improves several layout and occlusion metrics over the evaluated baselines while maintaining competitive text-image alignment.
Figures & tables
Figure 1: Our training-free framework resolves attention competition through attention arbitration, enabling flexible depth ordering, robust generation in complex occlusion scenarios, and multi-layer occlusion synthesis.
Figure 2: The overall architecture of DepthArb. The framework converts U-Net cross-attention or the image-to-text component of MMDiT joint attention into a unified spatial-text representation. Layout Confinement establishes spatial alignment, while Occlusion Conflict Estimation adaptively reweights the depth-aware orthogonality of Attention Arbitration Modulation and the compactness regularization of Spatial Compactness Control.
Figure 3: Qualitative comparisons on representative occlusion cases. Each row corresponds to one prompt with specified object ordering, and columns show the input layout, baselines, and DepthArb, respectively.
Method
Layout Alignment (%)
Text–Image Consistency
Occlusion Quality
Inference Time
mIoU-fg ↑
mIoU-bg ↑
mIoU-all ↑
CLIP Score ↑
FOCR (%) ↑
BOR ↑
FBS ↑
Time (s) ↓
SDXL
30.71
42.03
36.73
33.44
56.20
24.42
69.7
3.81
FLUX
36.98
50.54
42.48
34.01
61.22
24.91
71.5
6.57
Layout Guidance
27.32
43.57
34.59
32.00
52.30
25.37
58.6
4.05 (+130%)
R&B
55.15
59.12
57.88
31.60
70.27
25.86
73.3
26.89 (+684%)
VODiff
56.01
58.12
57.98
32.39
73.55
25.17
76.1
8.85 (+48%)
Table 1: Quantitative comparison of base models, training-free layout-guided baselines, four SDXL-based ablation variants, and the full DepthArb models on OcclBench. Percentage overheads are computed relative to each method’s corresponding base model. Bold and underline denote the best and the second best results, respectively.
Method
OverLayBench-Regular
OverLayBench-Complex
mIoU ↑
O-mIoU ↑
SR E ↑
SR R ↑
CLIP G↑
CLIP L↑
mIoU ↑
O-mIoU ↑
SR E ↑
SR R ↑
CLIP G↑
CLIP L↑
BoxDiff
19.40
5.33
37.58
71.81
36.50
19.97
20.02
5.21
33.52
76.41
36.92
19.91
Layout Guidance
15.81
4.51
44.94
72.76
31.76
20.05
16.34
4.01
37.76
75.53
32.75
19.72
R&B
20.35
5.54
32.85
65.01
34.49
19.88
19.80
4.85
28.38
69.97
34.57
19.47
VODiff
42.80
19.67
50.85
69.03
30.27
22.33
42.48
20.88
43.67
70.20
32.98
20.58
LaRender
57.21
30.06
54.97
75.40
28.67
26.07
52.27
25.07
45.08
74.62
26.85
24.43
Table 2: Comparison between training-free methods on OverLayBench ( Li et al. 2025 ) . Bold and underline denote the best and the second best results, respectively.
Figure 4: Visual ablation of component contributions. While the base model with only layout constraints exhibits concept mixing, the full DepthArb model produces clearer foreground–background separation in this example.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Figure A.1: Distribution of the number of occlusion pairs (top), object counts (bottom left), and depth hierarchy levels (bottom right).
Figure A.2: The prompt template used to automatically construct OcclBench. Given a candidate object pool sampled from MS-COCO, the LLM selects 2–6 object categories, synthesizes a semantically coherent scene, and generates geometric annotations including bounding boxes, continuous depth values, and pairwise occlusion relationships.
Benchmark
Scale
Objects
Depth Annotation
Complex Multi-Object
T2I-CompBench++
1000
2
None
✗
VOBench
200
2–5
Visibility Order
△
RealOcc
70
2–5
None
△
BindBench
200
3–5
Visibility Order
✓
OcclBench (Ours)
310
2–6
Continuous Depth
✓
Appendix
Table A.1: Comparison of OcclBench with existing benchmarks for occlusion-aware text-to-image generation.
Figure A.3: The detailed system prompt utilized for the automated Foreground-Background Separability (FBS) evaluation.
Figure C.1: Analysis of key hyperparameters. We analyze λortho (top) and λcompact (bottom), with red values indicating default settings. Excessive λortho disrupts background integrity, while extreme λcompact values lead to either blurry boundaries or unnatural background style collapse.
SDXL
FLUX
Layout Guidance
R&B
VODiff
LaRender
Ours+SDXL
Ours+FLUX
UniNet ↑
0.354
0.383
0.327
0.330
0.361
0.378
0.386
0.410
CLIP Score ↑
31.45
32.01
30.46
30.04
31.62
31.22
31.68
33.14
Appendix
Table D.1: Quantitative comparison on T2I-CompBench++. DepthArb with FLUX obtains the highest UniNet and CLIP scores among the evaluated methods. Bold and underline denote the best and the second best results, respectively.
Figure E.1: Qualitative results of DepthArb with SDXL on the T2I-CompBench++ benchmark.
Figure E.2: Qualitative results of DepthArb with FLUX on the T2I-CompBench++ benchmark.
Figure E.3: Qualitative results across diverse artistic styles. DepthArb follows the specified spatial layouts while preserving the intended styles in these examples.
Ours vs. SDXL
Ours vs. LaRender
Winning Rate (%)
71.3
73.1
Appendix
Table F.1: Pairwise user study results. Participants were asked to choose their preferred result between our method and each baseline independently.
Figure F.1: Example image pairs used in our user study. Participants compared DepthArb against SDXL (left) and LaRender (right). In these examples, DepthArb follows the specified spatial relationships more closely than the corresponding baseline.
Figure G.1: Representative failure cases of DepthArb.
We identify occlusion reasoning as a fundamental yet overlooked aspect for 3D layout-conditioned generation. It is essential for synthesizing partially occluded objects with depth-consistent geometry and scale. While existing methods can generate realistic scenes that follow input layouts, they often fail to model precise inter-object occlusions. We propose SeeThrough3D, a model for 3D layout conditioned generation that explicitly models occlusions. We introduce an occlusion-aware 3D scene representation (OSCR), where objects are depicted as translucent 3D boxes placed within a virtual environment and rendered from desired camera viewpoint. The transparency encodes hidden object regions, enabling the model to reason about occlusions, while the rendered viewpoint provides explicit camera control during generation. We condition a pretrained flow based text-to-image image generation model by introducing a set of visual tokens derived from our rendered 3D representation. Furthermore, we apply masked self-attention to accurately bind each object bounding box to its corresponding textual description, enabling accurate generation of multiple objects without object attribute mixing. To train the model, we construct a synthetic dataset with diverse multi-object scenes with strong inter-object occlusions. SeeThrough3D generalizes effectively to unseen object categories and enables precise 3D layout control with realistic occlusions and consistent camera control.
Text-to-image diffusion models fail to generate correct object counts in dense scenes, where overlapping instances collapse into indistinguishable structures despite appearing visually plausible. We identify this as instance ownership collapse: tokens from overlapping objects interact freely through attention, while heavily occluded instances receive weak supervision due to their small visible areas. We address this through layout-aware attention biases that softly bias token interactions toward region-consistent grouping and suppress cross-instance leakage, paired with an amodal-balanced loss that amplifies gradients for occluded objects based on their occlusion level. To enable systematic evaluation, we introduce OverlapDepth-45K, a benchmark of densely overlapping scenes with amodal supervision. Our approach substantially improves count accuracy and prevents instance merging while preserving image quality. Project page: https://bachngoh.github.io/AIBL
Bach-Hoang Ngo, Si-Tri Ngo, Hieu Le +1
University of Science, Ho Chi Minh City, Vietnam · 2Vietnam National University, Ho Chi Minh City, Vietnam · 3UNC Charlotte, North Carolina, United States
Recent layout-to-image models have achieved remarkable progress in spatial controllability. However, they still struggle with inter-object occlusion. When bounding boxes overlap, most existing methods lack explicit occlusion information, which makes the generation in intersection regions inherently ambiguous and hinders the determination of complex occlusion relationships. As a result, they often produce entangled textures or physically inconsistent layering in the overlapped areas. To address this issue, we first construct SA-Z, a large-scale dataset enriched with explicit occlusion ordering and pixel-level annotations. Building upon our proposed dataset, we introduce OcclusionFormer, a novel occlusion-aware Diffusion Transformer framework that explicitly models Z-order priority by decoupling instances and compositing them via volume rendering. Furthermore, to ensure fine-grained spatial precision, we introduce a queried alignment loss that explicitly supervises individual instances and enhances semantic consistency. The proposed method effectively reduces ambiguity in overlapping regions, enforces correct occlusion dependencies, and preserves structural integrity, leading to substantial accuracy gains across diverse scenes.
Ziye Li, Henghui Ding
Institute of Big Data, College of Computer Science and Artificial Intelligence, Fudan University, China.