Abstract

In this paper, we introduce DC (Decouple)-ControlNet, a highly flexible and precisely controllable framework for multi-condition image generation. The core concept of DC-ControlNet lies in decoupling the control conditions, breaking down global control into an organic combination of distinct elements, contents, and layouts. By enabling users to flexibly mix these individual conditions, the framework allows for greater freedom and precision in generating desired images. Current ControlNet models require the entire image condition to be provided, which presents two main challenges. First, global image conditions lack flexibility and are difficult to edit. Second, they cannot effectively handle the occlusion between different elements, often resulting in artifacts. With these challenges in mind, the proposed DC-ControlNet considers both intra-element attributes and inter-element interactions. For individual elements, we propose the Intra-Element Controller, which takes various types of conditions as content attributes and layout attributes. Furthermore, we introduce the Inter-Element Controller to address the challenges of multi-element interaction and occlusion. This module combines multiple elements according to user-defined logical relationships, ensuring coherent and accurate integration. Extensive evaluations demonstrate that DC-ControlNet significantly outperforms existing ControlNet models and Layout-to-image generative models in terms of single-element controllable generation flexibility and multi-element occlusion-aware generation.