InstanceAssemble was proposed in September 2025 by a research team from Fudan University and Xiaohongshu, and the relevant research results were published in a paper. InstanceAssemble: Layout-Aware Image Generation via Instance Assembling AttentionIt was selected for NeurIPS 2025.

InstanceAssemble is a novel approach for layout-to-image generation that sequentially processes global text hints and layout conditions, achieving robust handling of complex layouts through independent attention mechanisms. By integrating layout conditions through instance-assembled attention mechanisms, this framework enables bounding box (bbox)-based positional control and multimodal content control including text and additional visual content. This method achieves flexible adaptation to existing DiT-based T2I models through a lightweight LoRA module.