MultiCompose: Multi-Concept Personalized Composition with Per-Subject Attribute Binding

2026-08-04

AI Summary

This work addresses the challenge in text-to-image generation where co-occurring multiple subjects often suffer from identity confusion and attribute misalignment due to the absence of mechanisms that reliably bind each subject to its designated attributes. To resolve this, we propose MultiCompose, a framework that decouples concept personalization from multi-subject reasoning and integrates semantic-preserving regularization, a two-stage generation pipeline, and spatially exclusive attention masks to achieve precise subject layout and attribute composition. We further introduce MSP-Bench, the first evaluation benchmark tailored for multi-subject scenarios, which jointly assesses identity fidelity, attribute binding accuracy, and misalignment severity. Experiments demonstrate that MultiCompose significantly outperforms existing methods on both conventional metrics and MSP-Bench, with the new benchmark effectively uncovering failure modes overlooked by traditional evaluation protocols.

Abstract

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

Problem

• multi-concept personalization
• attribute binding
• identity degradation
• attribute misalignment
• text-to-image generation

Innovation

• multi-concept personalization
• attribute binding
• spatially exclusive composition
• diffusion models
• benchmark evaluation