Multimodal Large Language Models for Multi-Subject In-Context Image Generation

Multi-subject image generation aims to synthesize images that faithfully preserve the identities of multiple reference subjects while following textual instructions. However, existing methods often suffer from identity inconsistency and limited compositional control, as they rely on diffusion models to implicitly associate text prompts with reference images. In this work, we propose Hierarchical Concept-to-Appearance Guidance (CAG), a framework that provides explicit, structured supervision from high-level concepts to fine-grained appearances. At the conceptual level, we introduce a VAE dropout training strategy that randomly omits reference VAE features, encouraging the model to rely more on robust semantic signals from a Visual Language Model (VLM) and thereby promoting consistent concept-level generation in the absence of complete appearance cues. At the appearance level, we integrate the VLM-derived correspondences into a correspondence-aware masked attention module within the Diffusion Transformer (DiT). This module restricts each text token to attend only to its matched reference regions, ensuring precise attribute binding and reliable multi-subject composition. Extensive experiments demonstrate that our method achieves state-of-the-art performance on the multi-subject image generation, substantially improving prompt following and subject consistency.

Subject-driven image generation aims to synthesize new images that preserve the identity of the given subject while following textual instructions. Existing approaches often encode text and reference images separately. This limits cross-modal reasoning abilities and causes copy-paste artifacts. Recent frameworks that connect multimodal models and diffusion models improve instruction following, but largely overlook identity preservation. To address these limitations, we condition diffusion models on Multimodal Large Language Models (MLLMs) that jointly encode text and reference images, and augment it with VAE-based identity conditioning. A novel Dual Layer Aggregation (DLA) module is designed to aggregate multi-level MLLM features for optimal conditioning, and a multi-stage denoising strategy is applied to progressively balance the semantic information from MLLM and fine-detail identity from VAE during inference. Extensive experiments demonstrate that our approach harmonizes multimodal understanding with identity preservation, mitigates copy-paste issues, and achieves superior performance regarding human preference on subject-driven image generation. Our project website is available at https://zsh2000.github.io/squeeze-mllm-subject-gen/.

Diffusion-based text-to-image generation has advanced significantly, yet customizing scenes with multiple distinct subjects while maintaining fine-grained control over their interactions remains challenging. Existing methods often struggle to provide explicit user-defined control over the compositional structure and precise spatial relationships between subjects. To address this, we introduce MS-CustomNet, a novel framework for multi-subject customization. MS-CustomNet allows zero-shot integration of multiple user-provided objects and, crucially, empowers users to explicitly define these hierarchical arrangements and spatial placements within the generated image. Our approach ensures individual subject identity preservation while learning and enacting these user-specified inter-subject compositions. We also present the MSI dataset, derived from COCO, to facilitate training on such complex multi-subject compositions. MS-CustomNet offers enhanced, fine-grained control over multi-subject image generation. Our method achieves a DINO-I score of 0.61 for identity preservation and a YOLO-L score of 0.94 for positional control in multi-subject customization tasks, demonstrating its superior capability in generating high-fidelity images with precise, user-directed multi-subject compositions and spatial control.