MultiCrafter: High-Fidelity Multi-Subject Generation via Spatially Disentangled Attention and Identity-Aware Reinforcement Learning

This paper introduces MultiCrafter, a framework for multi-subject driven customized image generation.

🔆 Motivation

Existing In-Context Learning based methods suffer from two main issues:

1. Attribute Leakage: When generating multiple subjects, their distinct features (like facial characteristics or clothing) often blend and fuse, a problem the authors attribute to "attention bleeding" between subjects. This severely compromises the identity and fidelity of each individual subject.
2. Failure to Align with Human Preferences: Models trained with simple reconstruction objectives do not capture nuanced human preferences for aesthetic quality or precise alignment with text prompts.

😉 The MultiCrafter Solution

To tackle these challenges, MultiCrafter proposes three core innovations:

• Identity-Disentangled Attention Regularization: This technique uses explicit positional supervision to force the model to learn separate attention regions for each subject, effectively reducing attribute leakage and enhancing subject fidelity.
• Efficient Adaptive Expert Tuning: It incorporates a Mixture-of-Experts (MoE-LORA) architecture to increase the model's capacity to handle diverse subjects and spatial layouts without increasing inference overhead.
• Identity-Preserving Preference Optimization: The framework designs a novel online reinforcement learning stage to align outputs with human preferences. This features a "Multi-ID Alignment Reward" for accurate fidelity scoring and uses the stable Group Sequence Policy Optimization (GSPO) algorithm for training.

Project Page: https://wutao-cs.github.io/MultiCrafter/