Source-linked AI summary
Inkspire: Supporting Design Exploration with Generative AI through Analogical Sketching
David Chuan-En Lin, Hyeonsu B. Kang, Nikolas Martelaro, Aniket Kittur, Yan-Ying Chen, Matthew K. Hong
TL;DR
Designers struggle with abstract prompting and with fixation-inducing T2I workflows that make generated designs difficult to extend. Inkspire addresses these issues through analogical inspirations, iterative sketching, and a sketch-to-design-to-sketch loop; studies found wider exploration and more collaborative ideation than with ControlNet, while later-stage refinement remains a boundary.
Problem
T2I workflows can induce design fixation, while designers struggle to express abstract concepts through prompts and to build on generated designs.
Method
Inkspire combines analogical inspirations, iterative sketching, sketch scaffolds, and a sketch-to-design-to-sketch feedback loop for co-creative design exploration.
Results
Inkspire users explored a wider design space and adopted a more conceptually iterative and collaborative ideation process than ControlNet users.
Takeaways & Limitations
Analogies and sketch scaffolds helped designers expand abstract concepts into alternatives and guide AI toward novel design intentions through iterative turn-taking.
Takeaways & Limitations
Inkspire primarily supports early exploration and may be less suitable when designers have a clear concept and need detailed later-stage refinement.
Abstract
from arXiv · showhide
With recent advancements in the capabilities of Text-to-Image (T2I) AI models, product designers have begun experimenting with them in their work. However, T2I models struggle to interpret abstract language and the current user experience of T2I tools can induce design fixation rather than a more iterative, exploratory process. To address these challenges, we developed Inkspire, a sketch-driven tool that supports designers in prototyping product design concepts with analogical inspirations and a complete sketch-to-design-to-sketch feedback loop. To inform the design of Inkspire, we conducted an exchange session with designers and distilled design goals for improving T2I interactions. In a within-subjects study comparing Inkspire to ControlNet, we found that Inkspire supported designers with more inspiration and exploration of design ideas, and improved aspects of the co-creative process by allowing designers to effectively grasp the current state of the AI to guide it towards novel design intentions.
1 INTRODUCTION
T2I models offer new creative possibilities, but designers face fixation and difficulty expressing abstract ideas through prompts. Inkspire addresses these challenges with analogical inspiration, iterative sketching, and sketch-to-design feedback, and users found it more exploratory and co-creative than ControlNet.
- T2I models can accelerate visualizing ideas and generate serendipitous inspiration, encouraging designers to adopt them in creative work.
- Designers using generative AI may repeatedly make small prompt changes around an initial intention, limiting exploration of novel design ideas.
- A formative exchange with professional designers identified unnatural text prompting, weak results for abstract concepts, and difficulty building on overly complete generated designs.
- Inkspire combines analogical inspirations, iterative sketching, and low-resolution sketch scaffolds derived from AI designs to support fluid exploration and reduce fixation.
- Compared with ControlNet, Inkspire produced significantly more inspiration and exploration, more concept sketches, and a more co-creative interaction.
- A within-subjects study found significantly higher inspiration, exploration, and co-creation attributes for Inkspire than for the baseline.
2 RELATED WORK
Prior work explores generative and guided-sketching interfaces for expanding design ideas, but fixation, limited prompting, and high-fidelity outputs remain challenges. Inkspire extends analogy-driven design and sketch scaffolding into a closed-loop generative workflow.
- Human-AI design research finds that generative systems can increase idea generation while also producing design fixation and limited creative exploration.
- Visual controls and multimodal interfaces can support more creative exploration than text-only prompting, but high-fidelity outputs may still constrain designers.
- 2.2 Analogy-Driven Design: Analogy-driven design draws inspiration from known domains to find novel solutions in a target domain.
- 2.2 Analogy-Driven Design: Multimodal research extends analogy-driven design into visual retrieval and representation, while language models can help translate abstract concepts into physical forms.
- 2.3 Guided Sketching: Guided-sketching systems support skill building, reference, and creative exploration through visual feedback, retrieval, and sketch guidance.
- 2.3 Guided Sketching: Creative Sketching Partner inspires new designs through retrieved sketches, whereas Inkspire uses generative AI rather than an existing sketch database.
- 2.3 Guided Sketching: Inkspire converts high-fidelity GenAI designs into abstracted sketch scaffolds, supporting a closed loop for iterating on designs that otherwise appear too complete.
3 FORMATIVE SESSION WITH DESIGN PROFESSIONALS
A day-long exchange with seven automotive designers revealed that text prompting is unnatural, abstract concepts are difficult for T2I models to render inspiringly, and generated designs are hard to extend. These findings motivated sketch-based interaction, concrete analogies, and a stronger feedback loop.
- 3 FORMATIVE SESSION WITH DESIGN PROFESSIONALS: The formative session involved seven professional product designers from a large automotive company across multiple design disciplines.
- 3.1.1 Design Goal 1. Sketching as a Natural Method of Interaction.: Designers described prompting as unnatural and preferred sketching, often beginning with a simple line or silhouette.
- 3.1.1 Design Goal 1. Sketching as a Natural Method of Interaction.: Design Goal 1 was to let designers interact with AI through sketching while progressing from abstract lines toward complete sketches.
- 3.1.2 Design Goal 2. Visually-Concrete Inspirations.: Design briefs are often abstract, but T2I models commonly produce poor or generic results for abstract terms such as protectiveness.
- 3.1.2 Design Goal 2. Visually-Concrete Inspirations.: Design Goal 2 was to visualize abstract themes through visually concrete, diverse analogical inspirations.
- 3.1.2 Design Goal 2. Visually-Concrete Inspirations.: Designers found AI-generated designs difficult to iterate on because they looked too complete and encouraged choosing or discarding designs.
4 INKSPIRE
Inkspire combines analogical inspiration, sketch-guided generation, and design-to-sketch scaffolding into an iterative product-design workflow. Its components help users move from abstract concepts to designs, refine them through sketches, and explore variations.
- System overview: Inkspire’s workflow combines Sketch2Design for generating designs from sketches and analogies with Design2Sketch for converting generated designs into lower-fidelity sketch scaffolds.The two components form a sketch-to-design-to-sketch interaction loop.
- Sketch2Design: LLMs generate visually concrete analogical inspirations from abstract concepts across nature, architecture, and fashion.The method prompts for concrete objects rather than adjectives and can branch from selected inspirations.
- Limitations: The current implementation cannot revisit earlier inspirations, explore multiple branches in parallel, or connect generated inspirations to the user’s sketch.These constraints are identified as potential future-work areas.
- Sketch2Design: Users can begin with a single stroke, add strokes iteratively, and use remixing to generate more diverse designs from the same inspiration and sketch.Inkspire progressively supports sketch-guided generation as the drawing develops.
- Design2Sketch: Inkspire converts generated designs into transparent sketch scaffolds so users can draw from prior outputs without being overly fixated on photorealistic renders.The scaffold also helps users overcome a blank canvas by providing a reduced-fidelity underlay.
- Design2Sketch: Design2Sketch intersects semantic boundaries with soft edges to retain key structural lines while producing a natural sketch-like appearance.Semantic segmentation filters redundant texture-driven edges from the soft-edge extraction.
5 USER STUDY
The user study compared Inkspire with a similarly structured ControlNet baseline in a counterbalanced within-subjects design. Twelve participants completed product-design tasks and questionnaires, supplemented by interviews and free-response feedback.
- Study design: The within-subjects study compared Inkspire with a ControlNet baseline using a prompt box and sketching canvas with a similar interface layout.The study examined Inkspire’s fit with designers’ workflows and its response to design pain points.
- Participants: Twelve participants included six professional designers and six novices with moderate drawing experience.Professional and novice participants differed in reported product-design and drawing experience.
- Measures: Questionnaires assessed creativity through the Creativity Support Index and human-AI collaboration through measures including controllability, communication, harmony, partnership, attribution, and ownership.Interviews and free-response questionnaires provided qualitative evidence about components, workflow integration, and improvements.
- Procedure: Participants received informed consent and an overview of the study procedures and system components before completing the tasks.The overview covered analogy generation, sketching, and scaffolding interactions.
- Procedure: Participants completed one Inkspire task and one ControlNet task involving lamp or chair design, with tool and task order counterbalanced.The themes were serenity for the lamp and fluidity for the chair.
6.1 Creativity
Inkspire improved reported exploration and inspiration relative to ControlNet, while other measured creativity attributes did not significantly differ. Participants described analogy inspirations as useful for ideation and exploring varied design directions.
- Creativity Support Index: 5.83 versus 3.83: exploration was significantly higher with Inkspire than with the baseline, t(11)=3.94, p<0.01, r=0.77, d_s=1.13.The reported means were μ=5.83, σ=1.27 for Inkspire and μ=3.83, σ=1.64 for the baseline.
- Creativity Support Index: 5.92 versus 4.00: inspiration was significantly higher with Inkspire than with the baseline, t(11)=3.44, p<0.01, r=0.72, d_s=0.99.The reported means were μ=5.92, σ=1.24 for Inkspire and μ=4.00, σ=1.41 for the baseline.
- Qualitative feedback: Participants found analogy inspirations helpful for early-stage ideation and reported exploring varied forms, styles, patterns, proportions, and high-level design directions.Changing high-level concepts was described as producing substantially different looks and concepts.
- Other creativity attributes: Engagement, expressiveness, tool transparency, and effort/reward tradeoff did not significantly differ between Inkspire and ControlNet.The authors report that Inkspire did not improve these attributes but did not appear to degrade them.
6.2 Human-AI Collaboration
Inkspire supported stronger human-AI collaboration than ControlNet, with higher ratings for controllability, communication, partnership, and attribution. Its interaction pattern also involved more varied analogies, iterative sketching, and frequent design generation rather than prompt refinement and fixation.
- Collaboration ratings: Inkspire significantly improved controllability, communication, partnership, and attribution ratings over ControlNet.Participants reported no significant differences in harmony or ownership.
- Prompting behavior: Participants used fewer prompts with Inkspire and produced more semantically diverse prompts than with ControlNet.Inkspire averaged 8.50 prompts versus 11.9, with BERTScore similarity of 0.51 versus 0.76.
- Sketching behavior: Inkspire users created fewer total strokes but iterated with more time between strokes, while the lower sketching frequency was not significant.Inkspire averaged 17.3 strokes versus 59.8 for ControlNet; sketching frequency was 12.5 versus 20.1 strokes/min, p=0.12.
- Sketching behavior: Inkspire users reported less expensive sketch strokes despite sketching less, consistent with a more fluid collaborative sketching experience.Stroke-expense ratings were 6.08 for Inkspire versus 3.75 for ControlNet.
- Prompting behavior: Inkspire users generated more analogical inspirations and explored multiple source categories, especially Nature, Architecture, and Fashion.Participants explored 4.58 analogical inspirations on average; Architecture was the most common final choice, while Nature was explored most frequently on average.
- Overall usage patterns: Interaction logs showed Inkspire users beginning with analogical ideation, iterating through sketches, and generating more new designs than ControlNet users.ControlNet users more often crafted a whole sketch and handed it to the AI, whereas Inkspire supported back-and-forth exploration.
6.3 Final Design Quality
Inkspire received higher final design-quality ratings than ControlNet, although the difference was not statistically significant. Interaction examples indicated greater conceptual diversity in Inkspire’s design trajectories than in ControlNet’s.
- Design quality: 5.92 versus 4.58: Inkspire received higher final design-quality ratings than ControlNet, but the difference was not statistically significant.Ratings used a 7-point Likert scale; p=0.06.
- Design diversity: Inkspire produced more diverse design trajectories, while ControlNet outputs often remained on one conceptual track.Examples showed Inkspire designs derived from exploratory analogies and ControlNet designs refined through incremental prompt changes.
6.4 Designer Ratings of Usage Experience Satisfaction
Participants reported a significantly better usage experience with Inkspire than with ControlNet. Interaction logs suggest that low-cost experimentation and reduced emphasis on filler strokes contributed to this experience.
- Usage satisfaction: 6.08 versus 4.33: participants rated their usage experience significantly higher with Inkspire than with ControlNet.Ratings used a 7-point Likert scale, with p<0.01.
- Usage satisfaction: Inkspire supported low-cost experimentation by enabling abstract sketching focused on the big picture rather than successive filler strokes.In ControlNet, participants often drew large strokes before generating anything and spent more time between trials.
7 DISCUSSION
Inkspire combines analogical anchors, stroke-based generation, and sketch scaffolds to support broader, more iterative co-creation with generative AI. The discussion reports improved exploration and agency, while noting boundaries for parallel exploration, later-stage rendering, collective fixation, and empirical validation.
- Inkspire’s contribution: Inkspire recommends analogical anchors and converts high-fidelity renders into low-resolution sketch scaffolds for iterative design exploration.Designers can use analogies as inputs, generate images from sketches, and continue iterating through stroke-by-stroke interaction.
- Inkspire’s contribution: Inkspire helped participants explore a wider design space through a more conceptually iterative and collaborative ideation process than ControlNet.ControlNet users primarily crafted full sketches and made small prompt changes before rendering.
- Interpretation: Prompt- and sketch-focused interfaces can make designers feel that inputs must be relatively complete before conveying intent, encouraging minor prompt iteration.The discussion connects this affordance to the ControlNet interaction pattern and related prior findings.
- Interpretation: Inkspire’s problem prompt and analogical keywords let a single stroke carry expressive meaning while producing an output relevant to the designer’s goals.The system also supports a shared mental model of the designer’s and AI’s intended actions.
- Interpretation: Participants reported that Inkspire exposed alternative designs, increased agency and ownership, and reduced emphasis on prompt engineering.These observations were associated with the combined use of analogies, stroke-based interaction, and scaffolds.
- Limitations: Inkspire still explores a single thread rather than parallel ideas, limiting its support for parallel exploration documented in prior design research.The authors suggest generating multiple analogies and images at each sketch step as one possible direction.
- Limitations: Inkspire is primarily designed for early exploration and may be less suitable for later-stage rendering when designers have a clear concept to refine.The discussion suggests adaptive controls that shift between exploratory and detail-focused interaction.
- Limitations: The study’s within-subjects evaluation used a relatively small sample of 12 participants, motivating larger-scale studies and ablations.The authors also identify component-level ablations, including automatic per-stroke generation and sketch scaffolding, as future work.
8 CONCLUSION
Inkspire combines analogical inspiration and sketch scaffolding to help designers expand abstract concepts into alternatives through iterative designer–AI interaction. This process supports guiding AI toward novel design intentions and may help overcome design fixation.
- Inkspire supports product design prototyping through analogical inspirations and a sketch-to-design-to-sketch feedback loop.
- The research proposes design guidelines and future directions for sketch-driven tools that support co-creation with generative AI models.
- Analogy and sketch scaffolding let designers turn a single abstract concept and single pen stroke into multiple design alternatives.
- Iterative turn-taking helps designers guide the AI toward novel design intentions with the potential to overcome design fixation.