Li Yuheng ; Liu Haotian ; Wu Qingyang ; Mu Fangzhou ; Yang Jianwei ; Gao Jianfeng ; Li Chunyuan ; Lee Yong Jae

Abstract

Large-scale text-to-image diffusion models have made amazing advances.However, the status quo is to use text input alone, which can impedecontrollability. In this work, we propose GLIGEN, Grounded-Language-to-ImageGeneration, a novel approach that builds upon and extends the functionality ofexisting pre-trained text-to-image diffusion models by enabling them to also beconditioned on grounding inputs. To preserve the vast concept knowledge of thepre-trained model, we freeze all of its weights and inject the groundinginformation into new trainable layers via a gated mechanism. Our model achievesopen-world grounded text2img generation with caption and bounding box conditioninputs, and the grounding ability generalizes well to novel spatialconfigurations and concepts. GLIGEN's zero-shot performance on COCO and LVISoutperforms that of existing supervised layout-to-image baselines by a largemargin.