KnobGen: Controlling the Sophistication of Artwork in Sketch-Based Diffusion Models

Amin Karimi Monsefi, Pouyan Boreshnavard, Mengxi Zhou, Wei-Lun (Harry) Chao, Alper Yilmaz, Rajiv Ramnath

CVEU Workshop, CVPR 2025, Nashville, Tennessee, USA, 2025

Figure from KnobGen: Controlling the Sophistication of Artwork in Sketch-Based Diffusion Models

Sketch-guided diffusion models generally assume that the input sketch is drawn with skill and detail. Given a rough novice drawing, they either over-commit to the crude strokes or ignore the user’s intent entirely. KnobGen introduces a dual-pathway design that separates coarse semantic intent from fine-grained stroke detail, exposed to the user as a single continuous control. Turning that knob moves generation smoothly between loose interpretation of a quick doodle and faithful adherence to precise line art, letting one model serve users across the full range of drawing skill.

Paper Code