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

Researchers introduced a subject-driven image generation method that combines multimodal large language models with identity conditioning.

The approach uses joint text-image encoders, VAE-based identity cues, Dual Layer Aggregation, and staged denoising.

These design choices reduce copy-paste artifacts and improve both prompt following and subject identity preservation.

The work could make personalized image generation more reliable for people, objects, and other target subjects.