This paper investigates the advancement and execution of major text-to-image era models from 2000 to 2024. By testing these models beneath identical equipment and natural conditions, four key execution measurements are measured and compared. The ponder at that point presents a gathering learning approach that leverages the personal qualities of the best-performing models. Moreover, the investigation addresses a developing challenge within the field: the propensity of models to create excessively photorealistic yields. To handle this, a novel strategy is proposed where a demonstrator is trained to produce pictures in the interesting fashion of a particular craftsman. A reliable visual watermark is inserted in each yield to protect creation and authenticity, situating the show as an imaginative instead of simply imitative generator.
Sbera et al. (Thu,) studied this question.