Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DepthArb: Training-Free Depth-Arbitrated Generation for Occlusion-Robust Image Synthesis

Hongjin Niu, Jiahao Wang, Xirui Hu, Weizhan Zhang, Lan Ma, Yuan Gao

Abstract

Text-to-image diffusion models frequently exhibit deficiencies in synthesizing accurate occlusion relationships of multiple objects, particularly within dense overlapping regions. Existing training-free layout-guided methods predominantly rely on rigid spatial priors that remain agnostic to depth order, often resulting in concept mixing or illogical occlusion. To address these limitations, we propose DepthArb, a training-free framework that resolves occlusion ambiguities by arbitrating attention competition between interacting objects. Specifically, DepthArb employs two core mechanisms: Attention Arbitration Modulation (AAM), which enforces depth-ordered visibility by suppressing background activations in overlapping regions, and Spatial Compactness Control (SCC), which preserves structural integrity by curbing attention divergence. These mechanisms enable robust occlusion generation without model retraining. To systematically evaluate this capability, we propose OcclBench, a comprehensive benchmark designed to evaluate diverse occlusion scenarios. Extensive evaluations demonstrate that DepthArb consistently outperforms state-of-the-art baselines in both occlusion accuracy and visual fidelity. As a plug-and-play method, DepthArb seamlessly enhances the compositional capabilities of diffusion backbones, offering a novel perspective on spatial layering within generative models.

DepthArb: Training-Free Depth-Arbitrated Generation for Occlusion-Robust Image Synthesis

Abstract

Text-to-image diffusion models frequently exhibit deficiencies in synthesizing accurate occlusion relationships of multiple objects, particularly within dense overlapping regions. Existing training-free layout-guided methods predominantly rely on rigid spatial priors that remain agnostic to depth order, often resulting in concept mixing or illogical occlusion. To address these limitations, we propose DepthArb, a training-free framework that resolves occlusion ambiguities by arbitrating attention competition between interacting objects. Specifically, DepthArb employs two core mechanisms: Attention Arbitration Modulation (AAM), which enforces depth-ordered visibility by suppressing background activations in overlapping regions, and Spatial Compactness Control (SCC), which preserves structural integrity by curbing attention divergence. These mechanisms enable robust occlusion generation without model retraining. To systematically evaluate this capability, we propose OcclBench, a comprehensive benchmark designed to evaluate diverse occlusion scenarios. Extensive evaluations demonstrate that DepthArb consistently outperforms state-of-the-art baselines in both occlusion accuracy and visual fidelity. As a plug-and-play method, DepthArb seamlessly enhances the compositional capabilities of diffusion backbones, offering a novel perspective on spatial layering within generative models.
Paper Structure (34 sections, 14 equations, 11 figures, 4 tables)

This paper contains 34 sections, 14 equations, 11 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Training-Free Spatial Guidance
  4. Occlusion-Aware Synthesis
  5. Preliminaries
  6. Latent Diffusion Models and Cross-Attention.
  7. Spatial-Semantic Correspondence.
  8. Method
  9. Overview
  10. Layout Confinement via Attention Guidance
  11. Attention Arbitration Modulation
  12. Spatial Compactness Control
  13. Staged Latent Optimization Pipeline
  14. Experiments
  15. Implementation Details
  16. ...and 19 more sections

Figures (11)

  • Figure 1: DepthArb enables robust, training-free occlusion synthesis. Under overlapping layouts, (a) baseline methods suffer from spatial interference, causing artifacts like concept mixing and illogical occlusion. (b) DepthArb resolves these conflicts through explicit attention arbitration, ensuring precise visibility control and correct object depth ordering even in complex scenes.
  • Figure 2: The overall architecture of DepthArb. The pipeline consists of two main complementary components: Attention Arbitration Modulation and Spatial Compactness Control. These regularizations are applied through a two-stage gradient guidance strategy. Early denoising steps enforce strict spatial disentanglement, while later steps relax the orthogonality constraint to synthesize natural boundaries and coherent textures.
  • Figure 3: Qualitative comparisons on representative occlusion cases. Each row corresponds to one prompt with specified object ordering, and columns show the ground-truth layout, baselines and our DepthArb, respectively.
  • Figure 4: DepthArb translates explicit depth values into robust occlusion relationships. The hierarchical structures illustrate the intended foreground-background ordering, which is consistently preserved across diverse semantic categories and scale variations.
  • Figure 5: Visual ablation of component contributions. While the base model with only layout constraints suffers from concept mixing, our full framework (DepthArb) successfully isolates the foreground from the background, ensuring accurate depth perception.
  • ...and 6 more figures