Conceptual
Browse Instruct Curate

Training-Free Reference-Based Sketch Extraction via Mixture-of-Self-Attention

A training-free technique (MixSA) that extracts stylized sketches from images by editing a pretrained diffusion model's self-attention layers, substituting their keys and values with those of reference sketches to transfer brushstroke style, control texture density, and interpolate between styles without any dedicated network training.