PaperlayerView exact revision on arXiv

Source-linked AI summary

Less Is More: Balancing Positive and Negative Space in Visual Concept Blending

Shishi Xiao, Adam J. Coscia, David H. Laidlaw

arXiv:2609.00476v1cs.CVcs.HC

TL;DR

Computational visual blending has underexplored where and how concepts should be composed using positive and negative space. This paper introduces a composition-aware pipeline that combines geometric region selection, multimodal agent planning, diffusion initialization, and vector refinement. Across quantitative comparisons and a user study, the method improves expressive, recognizable, and creative blending while producing editable outputs and supporting multiple design applications.

  • Problem

    Prior visual-blending methods focus mainly on what to blend and pay less attention to spatial composition, leaving where and how concepts should be integrated underexplored.

  • Method

    The pipeline combines geometric constraints with vision-language planning for region selection, then uses hybrid pixel–vector generation and refinement for contour-level concept integration.

  • Results

    The method improves expressiveness, recognizability, and creativity over existing approaches in quantitative comparison and a human study, with compositional alignment reaching 0.902 versus 0.647 for the second-best baseline.

  • Takeaways & Limitations

    The approach produces coherent, editable vector blends and demonstrates application across controllable image and infographic generation and other design domains.

  • Takeaways & Limitations

    The evaluation used novice participants and does not establish how practicing graphic designers would use the system or whether its guidelines match professional practice.

Abstract

from arXiv · show

Graphic designers often blend visual concepts to communicate multiple ideas within a single image, leveraging positive and negative space to create balance, emphasis, and aesthetic appeal. While computational methods have begun to support automatic concept blending, they largely overlook the role of spatial composition in the design. To address this gap, we present an automatic pipeline that explicitly applies positive and negative space throughout the blending process. Our approach first identifies plausible regions for concept integration by combining semantic reasoning from vision-language models with geometric constraints derived from real-world examples. Conditioned on these regions, the system generates blended compositions using a hybrid pixel-vector pipeline: diffusion-based inpainting produces a fast, coarse initialization, which is then refined through vector-based optimization at the point level to ensure structural coherence and balanced semantic expression. A multimodal agent orchestrates this process as a planner and evaluator, enabling iterative improvement and interpretable control. Through an evaluation using both baseline comparisons and a user study, we demonstrate greater expressiveness, creativity, and concept recognizability by effectively leveraging positive and negative space. We further demonstrate the generalizability of our approach across diverse applications, including controllable image and infographic generation.

1 INTRODUCTION

Visual blending combines multiple concepts within a coherent form, but prior computational methods have underexplored where and how concepts should be spatially composed. The paper introduces an automatic pipeline that derives spatial guidelines and combines geometric reasoning, agent planning, and hybrid pixel–vector blending.

  • Motivation: Visual blending conveys dual meanings through a unified representation, often using shared contours and alternating positive and negative space.The Rubin vase illustrates how one contour can define both a vase and two faces.
  • Research gap: Prior automated methods emphasize what to blend while paying less attention to where and how concepts should be composed.The paper frames spatial composition as important for balancing expressiveness, creativity, and recognizability.
  • Research gap: The paper addresses two challenges: identifying blendable spatial regions without breaking structural coherence and constructing shared contours that preserve both concepts’ recognizability.These correspond to the questions of where to blend and how to blend.
  • Contributions: Its contributions are design guidelines derived from real-world examples and a general pipeline that supports automatic concept blending.The pipeline applies the guidelines computationally rather than limiting them to manual authoring.
  • Approach: The pipeline uses geometric candidate regions and multimodal planning, then combines diffusion inpainting with vector refinement for coherent, editable blends.The two stages sequentially address region selection and concept integration.
  • Evaluation: Quantitative comparison and a human study show improved expressiveness, recognizability, and creativity over existing approaches, with applications including controllable images and infographics.The evaluation also examines the framework’s versatility across applications.

2 RELATED WORK

Related work spans perceptual visual blending, conceptual combination, automated generative methods, interactive design systems, and domain-specific artifacts. These approaches provide semantic and generative capabilities, while the paper positions its contribution as a composition-aware, hybrid pixel–vector pipeline.

  • Perceptual foundations: Visual blending research includes reversible figures and hybrid images in which multiple interpretations vary with viewing conditions.These works establish perceptual ambiguity as a foundation for combining meanings in one representation.
  • Conceptual theories: Conceptual blending theories describe emergent meaning, whereas analogy and conceptual combination emphasize alignment and integration without requiring new meaning.The paper distinguishes these accounts from its computational visual-blending setting.
  • Generative methods: Diffusion-based methods automate multi-interpretation images through component conditioning, multi-view aggregation, and multi-condition optimization.Related applications include visual anagrams, multi-view illusions, and task-specific blends.
  • Interactive systems: HCI systems support metaphor generation, structured decomposition, conceptual exploration, and multimodal design workflows.Examples include VisiBlends, VisiFit, MetaMap, CreativeConnect, GraphicBench, and PosterCopilot.
  • Domain applications: Domain-specific systems embed semantics into typography, data visualizations, and QR codes while preserving task-relevant structure or readability.The paper contrasts these task-specific systems with its broader pipeline.
  • Positioning: Pixel-based methods offer generative flexibility, while vector-based methods provide explicit contour control; this paper combines both strengths.Differentiable rendering and diffusion priors support optimization of vector parameters.

3 DESIGN GUIDELINES

The design guidelines treat spatial composition and contour structure as central to visual blending. They distinguish positive–negative and positive–positive modes, establish interpretation hierarchy, and use geometric supports to locate contour regions for integration.

  • Empirical basis: The guidelines are derived from a formative analysis of 100 real-world graphic-design, logo, and visual-metaphor examples.The analysis identifies recurring concept roles, spatial compositions, and shared contours rather than claiming an exhaustive taxonomy.
  • Empirical basis: Blending examples are annotated by concepts, spatial roles, shared contours, and geometric guides.These annotations connect semantic interpretation with the shape structure used for computation.
  • Spatial modes: Positive–negative blending places one concept in the foreground and the other in negative space, whereas positive–positive blending expresses both within the foreground shape.Existing computational approaches have primarily focused on positive–positive blending.
  • Design goals: Effective blends establish an interpretation hierarchy in which the primary concept is immediately recognizable and the secondary concept emerges through localized cues.This hierarchy reflects the unequal perceptual emphasis typically assigned to blended concepts.
  • Contour optimization: Contour-level optimization adapts selected segments to encode the secondary concept while preserving the primary silhouette and structural coherence.Local contour segments can represent fine-grained components, whereas the overall contour carries global identity.
  • Design goals: The primary concept defines the global silhouette, while the secondary concept is integrated locally along selected contour regions.This asymmetric assignment preserves the primary shape’s overall reading while enabling secondary reinterpretation.
  • Region selection: Candidate regions occur where the primary shape deviates from geometric support: concave residuals for positive–negative blending and outward protrusions for positive–positive blending.The positive–negative case uses the convex hull, while the positive–positive case uses an interior support region.

4 METHOD

The method separates where-to-blend region selection from how-to-blend contour integration. A planner combines geometric and semantic reasoning, while a hybrid pixel–vector process initializes, evaluates, and refines the blended shape.

  • Input and representation: Users provide primary and secondary concepts plus a positive–positive or positive–negative blending type that controls their spatial alignment.The primary concept supplies the base geometry, while the secondary concept is injected during blending.
  • Input and representation: The primary concept is generated as a silhouette and converted into an SVG representation for vector-based region detection and shape refinement.The SVG uses parameterized Bézier curves to support iterative deformation control.
  • Stage I: region selection: Stage I detects structurally suitable contour segments with geometric guides and refines them using an AI planner that considers semantic compatibility.The output is a set of contour segments selected for secondary-concept integration.
  • Stage I: region selection: The planner agent consolidates candidate regions by combining geometric characteristics with the semantic roles of contour segments.This adds semantic prioritization to geometry-based candidate generation.
  • Stage II: pixel initialization: Stage II uses diffusion inpainting with a progressively widened mask to provide a fast semantic initialization in pixel space.A complementary forbidden mask limits edits from bleeding into the rest of the primary shape.
  • Stage II: vector refinement: The selected contour segment is locally refit to the inpainted region, preserving the rest of the contour and its established vector indexing.This transfers the pixel-space initialization back into the vector representation without re-vectorizing the whole shape.
  • Stage II: evaluation and optimization: An evaluator scores global and local concept expression, and low-scoring components are prioritized during differentiable vector optimization with diffusion-based semantic guidance.In practice, components with scores below τ=0.5 are flagged for refinement.
  • Outputs: The resulting vector outputs support stylized variants and remain fully editable for downstream design workflows.The pipeline demonstrates editable examples across its blending cases.

5 Evaluation

The evaluation combines quantitative baseline comparisons, a novice user study, and downstream application cases to assess structural integration, interpretation, and generalizability. Across these evaluations, the method improves compositional alignment and user-rated creativity and expressiveness while supporting controllable images and data-driven infographics.

  • Evaluation design: The evaluation combines quantitative comparison against commercial models, a 12-person user study, and application demonstrations.The studies assess interpretation of intended concepts, blending regions, and blending modes rather than optimizing a single automated metric.
  • Quantitative comparison: 39.4%: the method improves compositional alignment over the second-best baseline, scoring 0.902 versus 0.647.It ranks third in perceptual quality and second in recognizability, remaining close to the top baselines.
  • Quantitative comparison: The method integrates concepts structurally by encoding secondary concepts through negative space and contour deformation rather than simple overlay or insertion.Examples include carving a city into an excavator while preserving its silhouette and giving a turtle and leaf a unified contour.
  • User study: 41.5% and 61.1%: the method’s IoU exceeds vector-based baselines, indicating blending regions more consistent with users’ expectations.Recognition remains stable across blending types, including positive–negative blending, with a modest response-time cost.
  • User study: The method matches pixel-based baselines on recognizability and outperforms all baselines on creativity and expressiveness.Friedman tests found significant differences across methods for every metric and blending type, with p < 0.001.
  • User study: Participants described layered interpretation, emergent meanings, and negative-space placements that became reasonable in retrospect despite being harder to predict.These responses connect structural integration with compound meanings such as plastic pollution and with potential applications including fables.
  • Case studies: The pipeline generalizes to controllable image generation and data-driven infographics using both positive and negative space.In charts, positive blending forms building-like bars while negative space forms tree silhouettes, allowing one visualization to convey dual data-driven semantics.

6 Discussion

The discussion argues that spatial composition should be an explicit design control and that meaning-driven blending requires higher-level concept support. It also identifies limitations in agent understanding, evaluation scope, and available datasets and models for negative concepts.

  • 6.1 Design Implications: Spatial composition should be treated as a first-class design control because where a concept is integrated affects alignment with human expectations and perceived recognizability, creativity, and expressiveness.The authors argue that future systems should explicitly reason about where to blend rather than treating spatial placement as a generation byproduct.
  • 6.1 Design Implications: Meaning-driven blending requires higher-level support because the pipeline does not determine which concept pairs or regions will produce emergent meaning.That framing remains the user’s responsibility, limiting the system’s support for metaphorical concept blending.
  • 6.2 Toward Agents that Do More with Less: The authors identify a broader gap in training and evaluation: generative agents struggle to convey implicit concepts through minimal visual structure.They characterize this challenge as models struggling to generate “less” rather than continually adding elements.
  • 6.2 Toward Agents that Do More with Less: Progress toward agents that understand negative concepts is constrained by the lack of datasets, models, and shared guidelines for mapping implicit concepts to design.The proposed resources could support design-aware planning modules and evaluators for refining agent behavior with human designers.
  • 6.3 Limitations and Future Work: The evaluation leaves open how professional designers would use the system, whether its guidelines match practice, and how it performs in interactive design tasks.The authors also identify extending evaluation beyond two case studies and developing negative-concept resources as future directions.

7 CONCLUSION

The paper presents an automatic composition-aware pipeline that models where and how concepts are integrated within a shared contour. By combining geometric reasoning, agent-based planning, and hybrid pixel–vector optimization, it produces coherent, expressive, editable designs across multiple design domains.

  • 7 CONCLUSION: The pipeline explicitly models where and how concepts are integrated within a shared contour.This composition-aware formulation is the paper’s central framing for automatic visual concept blending.
  • 7 CONCLUSION: Geometric reasoning, agent-based planning, and hybrid pixel–vector optimization produce coherent, expressive, and editable vector designs.The conclusion links these components to improvements in expressiveness, recognizability, and creativity.
  • 7 CONCLUSION: The approach supports flexible applications across multiple design domains.The authors position the work as a basis for future tools for creative visual communication.

A Input Processing

The implementation represents each closed contour with 150 control points, forming an ordered point set along the contour boundary.

  • A Input Processing: 150 control points represent each closed contour in the implementation.The resulting contour is denoted as P = {p_i}^N.

B Stage II: Blending and Refinement

Stage II integrates a secondary concept into a selected primary contour using hybrid pixel-space generation and vector-level optimization. The blending mask is rasterized from the selected segment and progressively expanded during inpainting.

  • B Stage II: Blending and Refinement: Stage II combines pixel-space generation with vector-level optimization to integrate a secondary concept into a selected primary contour.The stage provides the mathematical and algorithmic details of this hybrid blending and refinement process.
  • B Stage II: Blending and Refinement: The blending mask is rasterized as a polyline and progressively widened across inpainting iterations to provide sufficient generation space.Mask expansion increases the polyline width during successive inpainting steps.

B.0.1 Concept Injection via Inpainting.

The method converts a selected raster-generated region into a vector-contour optimization problem. It extracts a target contour as a point cloud and rearranges Bézier control points within the selected segment.

  • The selected contour segment S is optimized to integrate rasterized details into the vector representation.
  • The method extracts the target contour from the inpainted region as a point cloud T.
  • Bézier control points within S are rearranged through constrained optimization.

B.0.2 Pixel–Vector Alignment.

The alignment stage matches the vector contour to generated content while regularizing its geometry and evaluating semantic expression during refinement.

  • The Chamfer loss aligns the vector contour with the target geometry.
  • A bending loss regularizes local curvature by penalizing sharp turns and preventing self-intersections or jagged artifacts.
  • An evaluator agent assesses the expressiveness of the in-progress blend to guide targeted refinement.

B.0.3 Blending Objectives Evaluation.

Blending objectives are evaluated through concept-specific expressiveness scores tied to contour segments, with optimization focused on under-expressed regions and exposed through an inspectable interface.

  • Each semantic component c_k is assigned an expressiveness score s_k and associated with a contour segment S_k.
  • Regions scoring below the τ=0.5 threshold are prioritized for refinement during contour optimization.
  • The prototype interface provides input specification, pipeline inspection, and stage-specific output inspection panels.
Loading 2609.00476v1…