Feature Fusion for Multi-Condition Controllable Image Generation

The rapid advancement of AI-generated content (AIGC) has led to methods that produce highly diverse images from text descriptions. However, relying solely on text descriptions often fails to meet the creator's precise needs. To address this issue, researchers have proposed various methods to enhance the controllability of generated images, which have achieved notable success. Nonetheless, a single condition often fails to fully express creative intentions. Combining multiple conditions with textual information can significantly enhance result controllability. This paper proposes a method for multi-condition controllable image generation based on feature fusion. It categorizes the given conditions into two types: details (e.g., pose maps, edge maps) and macro conditions (e.g., semantic segmentation maps, depth maps). These categories describe the image's detailed structure and background style, respectively. Leveraging their distinct characteristics and requirements, the method uses a combination of structural and global attention mechanisms for feature fusion. This approach extracts high-level and low-level conditional features, which are then fused into an image incorporating all given conditions using a feature pyramid. The composite image is fed into the control module of the generation model, guiding the generation process. Through experimental visualization and data metrics analysis, our method demonstrates improved controllability of generated results compared to existing approaches. It effectively satisfies multiple descriptive conditions while ensuring a certain level of quality and diversity in the generated images.