ريحانه نجف زاده اسكوئي

Abstract With the growing use of image content generation, there is an increasing need fo‎r generative models whose outputs are simultaneously high-quality an‎d diverse. Diffusion models, which reverse a gradual noising process, are capable of producing high-quality images, but they typically struggle to balance quality an‎d diversity. Strengthening conditioning o‎r reducing noise in the input improves quality but limits diversity, an‎d the opposite is also true. This study introduces a method called “multi-reference guidance,” in which two different reference examples are selec‎ted an‎d used fo‎r guiding the diffusion process. To control this balance precisely, the guidance is applied dynamically an‎d dependently, similar to classifier-free guidance, an‎d the publication step is adjusted. Sampling is perfo‎rmed in a cold-diffusion space. Experimental results show that the proposed method increases output diversity without noticeable quality loss. Considering limited computational resources in this study, further improvement is expected at larger scales.

مدل‌هاي انتشار , مبدل , هوش مصنوعي , نويز پويا , هدايت محتوايي , تنوع خروجي , راهنماي چند مرجعي

Diffusion models , Transformer , Artificial intelligence , Dynamic noise , Content-guided , Diversity , Multi‑reference guidance

Reyhaneh Najafzadeh Oskuee

Dr.Farzan Haddadi