Core Concepts
Disentangling similarity from controllability in text-to-image customization for optimal results.
Abstract
Text-to-image customization revolutionizes content creation.
Existing pseudo-word paradigm faces dual-optimum paradox.
RealCustom disentangles similarity from controllability.
"Train-inference" framework for real-time open-domain customization.
Adaptive scoring module and mask guidance strategy for precise customization.
Superior similarity and controllability demonstrated in experiments.
RealCustom enables high-quality customization in real-time open-domain scenarios.
Stats
"RealCustom achieves 8.1% improvement in CLIP-T and 223.5% improvement in ImageReward for controllability."
"RealCustom achieves state-of-the-art performance in CLIP-I and DINO-I for similarity."
Quotes
"RealCustom achieves the unity of high-quality similarity and controllability in the real-time open-domain scenario."