Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces MultiCompose, a novel framework for generating personalized images with multiple subjects, each bound to specific attributes like clothing and accessories. By employing a semantic preservation regularization and a two-phase inference procedure, MultiCompose effectively decouples concept personalization from multi-subject inference, addressing issues of identity degradation and attribute misalignment that plague existing methods. The authors also present MSP-Bench, a new benchmark that evaluates identity fidelity, attribute binding accuracy, and attribute misalignment, demonstrating that MultiCompose significantly outperforms prior techniques on both traditional metrics and this new benchmark.
MultiCompose reveals how to maintain distinct identities and accurately bind attributes in multi-subject image generation, outperforming existing methods and exposing previously hidden failure modes.
Text-to-image diffusion models enable personalization of specific visual concepts from a small number of reference images. However, generating a single image that contains multiple personalized subjects, each bound to user-specified attributes such as clothing, accessories, and held objects, remains largely unaddressed. Without explicit spatial constraints, concurrently activated concept checkpoints produce overlapping cross-attention responses, causing per-subject identity degradation and attribute misalignment. Moreover, no established benchmark jointly evaluates these two failure modes in the personalized multi-subject setting. We present MultiCompose, a composition framework that decouples per-concept personalization from multi-subject inference. A semantic preservation regularization maintains attribute binding capacity during fine-tuning, while a two-phase inference procedure automatically establishes subject layout and composes per-concept predictions through spatially exclusive masks. We further introduce MSP-Bench, a benchmark that jointly evaluates identity fidelity (ID), attribute binding accuracy (BIND), and attribute misalignment (MIS) through a dual-pathway protocol. Experiments show that MultiCompose outperforms existing methods on both conventional metrics and MSP-Bench, confirming the benchmark's ability to reveal failure modes that conventional metrics overlook. Code is available at https://github.com/I2-Multimedia-Lab/MultiCompose