Squeezing Capacity from Multimodal Large Language Models for Subject-driven Generation

Researchers propose a subject-driven image generation framework that combines multimodal large language models with reference images to jointly encode text and visual cues.

The method adds VAE-based identity conditioning and a Dual Layer Aggregation module with multi-stage denoising.

This helps diffusion models follow text instructions while preserving subject identity more faithfully.

The approach reduces copy-paste artifacts and improves overall image quality.