Less Redraw, More Explore: Suggestion and Completion for Sketch-to-Image

Paper Title

Less Redraw, More Explore: Suggestion and Completion for Sketch-to-Image

Publication Info

• Topic area: Sketch-to-image systems and co-creative AI interaction techniques.
• Keywords: Sketch-to-image, creativity support, iterative design, co-creative AI, AutoSketch, BackSketch, user agency, exploration, expressiveness, human-AI collaboration.

Background and Problem

• Problem / challenge: Current sketch-to-image systems rely on static input-output workflows, forcing users into tedious redraw-regenerate cycles that hinder iterative creativity and exploration.
• Significance: Reducing iteration friction in sketch-to-image workflows can enhance creative exploration and expressiveness, making these systems more accessible and effective for casual users.
• Motivation and related work: Prior research has focused on algorithmic advances and interactive interfaces but has largely treated sketches as static inputs, offering limited support for iterative refinement. This paper addresses the gap by introducing dynamic, reversible interaction techniques.

Solution

• Proposed approach: Two novel interaction techniques—AutoSketch (pre-generation sketch completion) and BackSketch (post-generation sketch suggestion)—to enhance sketch-to-image workflows.
• Novelty:
  1. Introduction of pre-generation completion (AutoSketch) and post-generation suggestion (BackSketch) techniques to reduce iteration cost and expand creative possibilities.
  2. Design insights for embedding iterative support into sketch-to-image pipelines, emphasizing reversible AI contributions to balance user agency and system initiative.
  3. Empirical evidence from a 30-participant study demonstrating improved exploration, expressiveness, and perceived co-creative partnership compared to a baseline system.
• Procedure and key techniques:
  • AutoSketch: Allows users to draw partial sketches and invoke AI-driven completions that add meaningful elements. Completions can be accepted, edited, or undone, enabling progressive refinement.
  • BackSketch: Converts generated images into multiple simplified sketches at different abstraction levels, offering users alternative sketches for further editing and exploration.

Results

• Concrete findings:
  • AutoSketch and BackSketch scored higher than the baseline on exploration (Q3: p = 0.014, p = 0.010) and expressiveness (Q4: p = 0.013, p = 0.056).
  • AutoSketch increased perceived agency compared to the baseline (Q7: p = 0.035).
  • Participants actively adopted new features, invoking AutoSketch completions 1.9 times per idea and selecting BackSketch suggestions 1.57 times per idea.
• Advantage over baselines: Both systems enhanced exploration and expressiveness, with AutoSketch also boosting agency. Participants preferred AutoSketch (57%) and BackSketch (33%) over the baseline (10%).
• Experiments / evaluation:
  • Controlled study with 30 non-expert participants using all three systems in counterbalanced order.
  • Measures included Likert-scale questionnaires (Q1–Q10), behavioral logs, and qualitative feedback.
  • Participants completed four-minute sketching sessions with at least two distinct sketches per system.
• Limitations and future work: Findings are limited to short-term, exploratory sessions with non-expert users. Future research should include longitudinal studies, professional workflows, and multimodal sketch-text pipelines.

Summary

This paper introduces AutoSketch and BackSketch, two novel interaction techniques for sketch-to-image systems that reduce iteration friction and enhance creativity support. AutoSketch provides pre-generation sketch completions, while BackSketch offers post-generation sketch suggestions, enabling fluid, reversible workflows. A user study with 30 participants demonstrated improved exploration, expressiveness, and agency compared to a baseline system. These findings highlight the importance of timing, reversibility, and constrained system initiative in co-creative sketch-based tools, with implications for broader applications in creative domains such as animation and conceptual illustration.