While diffusion models have revolutionized the field of generative artificial intelligence, their application in generating complex human sketches remains challenging. Current vector-based generation paradigms struggle to handle high-complexity sketch datas due to structural limitations in data representation, while direct application of diffusion models often results in insufficient sketch recognizability or “over-sketching”artifacts caused by the abstract and sparse nature of sketches. This study systematically analyzes the root causes of these issues, revealing limitations in existing methods for determining optimal guidance scales and incompatibilities in optimization objectives triggered by conventional training mechanisms. To address these challenges, we propose: i) A Scale-Adaptive Guidance strategy that dynamically optimizes classifier guidance scale parameters to establish an adaptive balance mechanism between recognizability and generation complexity; ii) A Classifier Representation Enhancement strategy that refines classifier training processes to ensure alignment between classifier objectives and diffusion model goals; iii) A Three-Phase Sampling strategy introduced during inference to enhance sketch diversity and generation quality. Experimental validation on the QuickDraw dataset demonstrates that our approach overcomes the limitations of traditional vector-based methods in complex sketch generation, showcasing the unique potential of diffusion models in this domain. The project code is available at: https: //github. com/HuJijin/AdaptiveGuidedSketch.
Li et al. (Sat,) studied this question.