multimodal llm

MLLMs for cleaner subject-driven image generation

This paper uses MLLMs plus VAE identity conditioning to improve subject-driven image generation and reduce copy-paste artifacts.