Reproduce “InstanceAssemble: Layout-Aware Image Generation via Instance Assembling Attention” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qiang Xiang, Shuang Sun, Binglei Li, Dejia Song, Huaxia Li, Nemo Chen, Xu Tang, Yao Hu and 1 more
A new framework for layout-to-image generation adapts Multimodal Diffusion Transformers using a lightweight LoRA-based module to achieve precise, instance-level control over object placement and content. This approach leverages a cascaded, independent attention mechanism to yield an mIoU of 94.97 on sparse layouts and 52.07 on dense layouts, while supporting multimodal input with minimal added parameters.
Key insights
- Processing each instance with independent attention on cropped image regions significantly improves spatial precision and semantic consistency in complex layouts.
- A cascaded design, which first processes global prompts and then refines with instance-wise conditions, helps maintain overall image quality while achieving layout adherence.
- Lightweight LoRA-based adaptation is effective for integrating layout control into large pre-trained Diffusion Transformers without sacrificing their generative capabilities or incurring high training costs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.