Mobile QR Code
Title Sketch-based Image Generation with ControlNet Feature Refinement through a Pretrained Diffusion Model Decoder
Authors 강민수(Minsu Kang) ; 송병철(Byung Cheol Song)
DOI https://doi.org/10.5573/ieie.2025.62.11.75
Page pp.75-78
ISSN 2287-5026
Keywords Diffusion model; Sketch-based image generation; Conditional control
Abstract Recent advances in diffusion models have enabled user-controllable image generation using text or image-based conditions. However, when using sketches as conditions, performance degrades due to ambiguous strokes and limited paired data. In this work, we propose a method that refines incomplete sketch control signals by leveraging the prior knowledge of pretrained diffusion models. Our approach achieves up to 18.1% improvement over existing methods in metrics such as FID-I, CLIP score, and LPIPS, demonstrating superior performance both quantitatively and qualitatively.