Presentation

· Contributors · Organizations · Search Program

Separate-and-Enhance: Compositional Finetuning for Text-to-image Diffusion Models

SessionControllable Image Generation and Completion

DescriptionThis work targets on improving the compositional capability of text-to-image models. Different from previous approaches that requires heavy test-time adaptation per prompt, we propose a compositional finetuning framework with two novel objectives. Through comprehensive evaluations, our model demonstrates superior performance in image realism, text-image alignment, and adaptability to novel concepts.

Authors

Zhipeng Bao

Carnegie Mellon University

University of Illinois Urbana-Champaign

Martial Hebert

Carnegie Mellon University