HistoryPalette: Supporting Exploration and Reuse of Past Alternatives in Image Generation and Editing

Paper Title

HistoryPalette: Supporting Exploration and Reuse of Past Alternatives in Image Generation and Editing

Publication Info

• Topic area: Enhancing generative image editing workflows through history exploration and reuse.
• Keywords: generative AI, image editing, history reuse, creative workflows, semantic organization, collaboration, inpainting, prompt-based generation, design alternatives, user studies.

Background and Problem

• Problem / challenge: Current generative AI tools lack effective mechanisms for exploring and reusing past design alternatives, forcing creators to rely on error-prone manual systems like undo/redo logs or saving separate files.
• Significance: Efficient reuse of prior alternatives can save time, reduce costs, and enhance creative workflows, especially in iterative design tasks.
• Motivation and related work: Prior research has focused on improving prompt refinement and chronological history navigation but has not addressed semantic organization or reuse of alternatives in pixel-based editing workflows. This paper seeks to fill this gap by leveraging semantic information from prompt-based workflows.

Solution

• Proposed approach: HistoryPalette, a system for exploring and reusing prior design alternatives in generative image editing workflows.
• Novelty:
  1. Introduction of three semantic history palettes (Position Palette, Concept Palette, Time Palette) for organizing and accessing alternatives.
  2. Filtering mechanism to remove unsuccessful generations using a vision-language model.
  3. Support for collaborative workflows by enabling shared exploration of design histories.
  4. Integration of inpainting and rasterization for seamless reuse of alternatives.
• Procedure and key techniques:
  • Alternatives are organized by spatial position, semantic concepts, and creation time.
  • Users can preview alternatives by hovering, reuse them via paste or rasterization, and filter out failed generations.
  • A clustering algorithm groups alternatives into concept categories, while positional encoding maps alternatives to specific regions of the canvas.
  • A version timeline provides access to full project versions.

Results

• Concrete findings:
  • Filtering achieved 93% accuracy (86% precision, 76% recall) in identifying unsuccessful generations.
  • Creative professionals reused prior alternatives 38.8% of the time, while client collaborators reused them 75.6% of the time.
  • Generation delays averaged 29.51 seconds, while rasterization delays averaged 12.89 seconds.
• Advantage over baselines:
  • Semantic organization (position and concept palettes) was preferred over traditional chronological organization (time palette).
  • Participants found reuse faster and more predictable than generating new alternatives.
• Experiments / evaluation:
  • Two user studies: one with three creative professionals (41 years combined experience) and one with eight client collaborators.
  • Metrics included interaction frequency, reuse rates, and qualitative feedback.
  • Creative professionals worked on open-ended projects, while client collaborators edited pre-populated compositions.
• Limitations and future work:
  • Findings are specific to experienced designers and open-ended workflows; further validation is needed for constrained domains and novice users.
  • Current limitations include boundary artifacts when reusing alternatives and challenges in scaling to longer histories.
  • Future work includes supporting direct manipulation workflows, dependency-aware reuse, and cross-project palette sharing.

Summary

HistoryPalette is a system designed to enhance generative image editing workflows by enabling the exploration and reuse of past design alternatives through semantic organization. User studies demonstrated that creative professionals and client collaborators preferred HistoryPalette's position and concept palettes over traditional time-based organization, finding them useful for improving efficiency, control, and collaboration. The system's filtering and clustering mechanisms further streamlined history management. While promising for open-ended creative tasks, future work is needed to address scalability, direct manipulation workflows, and broader applicability across domains.