OPTIMIZATION OF NEURAL NETWORKS IN DIFFUSION MODELS FOR SKETCH GENERATION USING BEZIER CURVE

Abstract

This article presents a comprehensive study regarding an innovative approach to optimising the computational process of generating artistic sketches, specifically focusing on the significant task of automating digital drawing workflows. Following a detailed and rigorous analysis of existing methodologies, a strategic decision was made to utilize parametric curve-based abstract data structures – specifically Bézier curves – as a superior alternative to the traditional, widely utilized raster pixel representations. This fundamental transition to using mathematical curve descriptions provides significantly broader opportunities for constructing compact, efficient models that yield higher quality and fully scalable generation results. Through this research, a robust model capable of generating novel geometric data from an input sketch was obtained, and the stability, consistency, and variability of the generation process were rigorously analyzed to ensure reliability. A critical component of this study involves a comparative analysis of the neural network parameters required for generating graphics in standard pixel formats versus the proposed Bézier curve method. The empirical results demonstrated that the employment of Bézier curves significantly reduces both the number of necessary input parameters and the overall computational complexity of the neural network architecture. To illustrate this computational disparity, creating a model to generate a standard 32 × 32 pixel image requires 1,024 input parameters (if the pixels are represented in black and white format) or 3,072 parameters (if in RGB format). In stark contrast, the model utilized to generate scalable curves requires only 800 parameters, representing a substantial reduction in computational load while maintaining structural integrity. The pivotal advantage of this new approach lies in its provision of perfect scalability; unlike traditional pixel (raster) formats, the generated data can be effortlessly enlarged or reduced to any dimension with zero loss of visual fidelity. Consequently, the quality of the generated image is no longer bound by the fixed resolution limitations inherent in raster formats but is determined solely by the potential and sophistication of the neural network architecture.