: ..
:
: 118
: ,
: 2025
: .. // . - 2025. - . 118. - .407-426.
: , - ,
(.): image generation, diffusion probabilistic models, generative models
: , . . , , , . PartiPrompts , Stable Diffusion v2.0. CLIPScore, ImageReward BRISQUE, . , . , .
(.): Currently, the leading approach for text-to-image generation is based on diffusion probabilistic models, as they enable the generation of high-quality and realistic images. A key property of diffusion models is their ability to produce diverse images from the same text prompt by utilizing different initial noise seeds. This paper identifies statistically significant patterns and correlations between the choice of initial noise, the category and complexity of the text prompt, on one hand, and the final image quality metrics, on the other. To achieve this, computational experiments were conducted to generate images based on prompts from the PartiPrompts dataset using different initial noise values. The generated images were produced using the Stable Diffusion v2.0 model and evaluated using the CLIPScore, ImageReward, and BRISQUE metrics, which reflect different aspects of image perception and alignment with the original prompts. The results demonstrate a strong and reproducible influence of specific initial noise seeds on all image quality metrics. This influence persists across prompts with varying semantics. The conducted analysis serves as a foundation for developing methods to assess and select the optimal initial noise for generating high-quality images using diffusion models.
PDF
: 37, : 22, : 18.