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

TL;DR

DepthArb introduces a training-free, attention-based framework that improves occlusion handling in text-to-image diffusion models, enhancing the realism and accuracy of overlapping object synthesis without retraining.

Contribution

It proposes DepthArb, a novel attention arbitration method with two core mechanisms, enabling better occlusion reasoning in diffusion models without additional training.

Findings

Outperforms state-of-the-art baselines in occlusion accuracy

Enhances visual fidelity in overlapping object synthesis

Provides a plug-and-play solution for diffusion models

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.