Design-Exploration Model Framework

Updated 14 July 2026

The Design-Exploration Model is a formal framework that structures early-stage design exploration by reifying intent, suggestions, and outcomes into modular blocks.
It introduces key constructs such as Exploratory Intent and Exploration Block to articulate open-ended design goals, control typicality, and enable non-linear branching via a provenance graph.
Empirical evaluations show that this model enhances exploration diversity, increases output iterations, and ensures continuity through context-aware chaining, branching, and reuse.

The Design-Exploration Model, in the sense distilled from "IdeaBlocks: Expressing and Reusing Exploratory Intents for Design Exploration with Generative AI," is a formal and interactional framework for early-stage design exploration with generative models. It treats exploratory intent as an explicit, structured object; packages that intent and its outcomes into reusable modules; and organizes exploration as a non-linear graph rather than a linear prompt history. In this formulation, design exploration is not only image generation, but the coordinated expression of what to explore, how broadly to explore it, how to preserve continuity across iterations, and how to revisit or adapt prior directions without restarting from scratch (Choi et al., 29 Jul 2025).

1. Problem setting and scope

The model arises from three limitations identified in a formative study with designers. First, there is difficulty expressing open-ended exploratory intent: designers often begin with broad, vague goals and struggle to articulate visual or abstract directions through text prompts, especially for properties such as style, pose, or atmosphere. They also lack a way to specify “how broad” they want the exploration to be. Second, there is a lack of continuity: when designers try to preserve successful aspects while changing other aspects, prompt reuse and inpainting can alter previously satisfying components unexpectedly or make them hard to localize. Third, there is limited reuse and iteration: typical tools provide only linear histories of prompts and outputs, whereas designers need both globally persistent preferences and locally scoped intents for branches (Choi et al., 29 Jul 2025).

Within this framing, the model targets broad exploration in the early phase of design rather than fine-grained convergence. It is therefore oriented toward divergent exploration, continuity across partial concept trajectories, and later revisitation. This differs from systems in which design space exploration is framed primarily as numerical optimization over objective functions, constraints, or Pareto fronts, as in "Practical Design Space Exploration" (Nardi et al., 2018) and "CAD Tool Design Space Exploration via Bayesian Optimization" (Ma et al., 2019). In the generative-design literature, by contrast, the emphasis is on sensemaking, branching, and the management of exploratory state, as also seen in DreamSheets, Inkspire, Expandora, DynEx, and StoryEnsemble (Almeda et al., 2023, Lin et al., 30 Jan 2025, Choi et al., 2 Mar 2025, Ma et al., 2024, Suh et al., 5 Aug 2025).

A common misconception is that exploration with generative AI is equivalent to repeated prompt editing. The model directly rejects that assumption by formalizing intent, provenance, and reuse as first-class entities. This suggests that the principal difficulty is not merely generating alternatives, but maintaining a manipulable representation of why those alternatives were produced and how they relate.

2. Core constructs and formal schema

The model is built around two formal constructs: Exploratory Intent and Exploration Block. Exploratory Intent is an explicit, early-stage design intent guiding broad exploration. It comprises three components: area of interest, direction, and range. The formalization is:

$I = (p, d, r),$

where $p \in P$ is the property being explored, $d \in \Sigma^*$ is a short direction string, and $r \in \{1,2,3,4,5\}$ is a typicality level controlling divergence, with $1 = \text{conventional}$ and $5 = \text{atypical}$ (Choi et al., 29 Jul 2025).

An Exploration Block is the modular unit that captures an Exploratory Intent and its outcomes for a property. Its schema is:

$b = (I, S, O, meta),$

where $I=(p,d,r)$ , $S=\{s_1,\dots,s_4\}$ are suggestions representative of the intended direction and range, $O=\{o_1,\dots,o_m\}$ are generated outputs tied to suggestions, and meta includes IDs, timestamps, lineage, and versioning. Blocks exist for both text-based properties such as entity, pose, and background concept, and image-based properties such as style, color palette, and lighting (Choi et al., 29 Jul 2025).

Construct	Formalization	Role
Exploratory Intent	$p \in P$ 0	Encodes property, direction, and typicality
Exploration Block	$p \in P$ 1	Stores intent, suggestions, outputs, and provenance
Exploration Graph	$p \in P$ 2	Organizes non-linear exploration across blocks

This schema matters because it separates three things that are often conflated in conventional prompt-based systems: the property under exploration, the semantic direction of change, and the desired breadth of variation. In the model, typicality is not a side effect of prompt wording but an explicit parameter. Suggestions are not final outputs; they are previews that support clearer intent articulation before generation. A plausible implication is that this intermediate representation reduces ambiguity by turning abstract, under-specified prompting into a structured decision about property, direction, and range.

3. Graph structure, continuity, and reuse

Exploration is represented as a non-linear graph:

$p \in P$ 3

where $p \in P$ 4 is the set of blocks and $p \in P$ 5 contains typed edges. Chain edges preserve context from upstream blocks. Reuse edges track provenance when blocks or paths are reused or adapted. Reference edges capture explicit references noted by the user. On the canvas, blocks are nodes and edges represent chaining and reuse; the result is a graph rather than a prompt log (Choi et al., 29 Jul 2025).

The graph supports four characteristic operations. Chaining links a new block to prior blocks so that the most recent image and prior descriptions become inputs to prompt synthesis for the next block. Branching creates parallel paths that explore alternative properties and directions. Revisiting reinserts previously stored blocks from a sidebar or copies entire paths, either literally or in context-adaptive form. Merging is not a distinct UI operation, but occurs when multiple upstream blocks are chained into one downstream block that integrates their contexts (Choi et al., 29 Jul 2025).

Continuity is maintained through three mechanisms. The first is context-aware prompt synthesis: before generating, the system extracts descriptions from the most recent upstream image and integrates them with the current block’s suggestion in order to preserve previously explored qualities. The second is provenance: each block persists with settings, suggestions, outputs, and lineage links, while the sidebar organizes property-wise histories and evolving directions. The third is versioning: copied blocks may preserve original IDs and settings in literal copies or produce adapted variants recorded as new versions connected by reuse edges (Choi et al., 29 Jul 2025).

These mechanisms make continuity a structural property of the exploration process rather than a fragile by-product of prompt reuse. Comparable concerns appear elsewhere in generative-design systems. DreamSheets emphasizes simultaneous comparison, duplication, and prompt provenance in spreadsheet form (Almeda et al., 2023); Expandora represents prompt evolution as a branching graph (Choi et al., 2 Mar 2025); StoryEnsemble uses node-link propagation across design stages (Suh et al., 5 Aug 2025). The distinctive move here is that the reusable object is not just an output or a prompt fragment, but the exploratory intent itself.

4. Algorithms and operational procedures

The model includes explicit procedures for property generation, suggestion diversification, image generation, and reuse. Given a design topic, GPT-4o suggests 8 relevant properties. For turning an Exploratory Intent into suggestions, the system uses a two-step LLM diversification pipeline. It first generates 10 alternative directions spanning typical to atypical variations for input direction $p \in P$ 6. It then produces multiple candidate suggestions for each direction: 10 per direction for text-based properties and 5 per direction for image-based properties, yielding candidate pools of 100 text suggestions or 50 image suggestions (Choi et al., 29 Jul 2025).

Typicality filtering is then applied. For text, alignment $p \in P$ 7 is computed via co-occurrence-based similarity using GloVe. For images, alignment is computed via CLIP score between the input direction and image candidates. Candidates are ranked by $p \in P$ 8, bucketed by alignment, and the bucket corresponding to the selected typicality level $p \in P$ 9 is chosen. Representative selection follows by clustering filtered candidates with $d \in \Sigma^*$ 0, using Sentence-BERT embeddings for text and CLIP visual embeddings (ViT-B/32) for images; cluster centroids become the four suggestions $d \in \Sigma^*$ 1 (Choi et al., 29 Jul 2025).

Generation after chaining uses a context-aware procedure. Step 1 extracts property descriptions from the most recent upstream image using GPT-4o. Step 2 retrieves the current block’s suggestion. Step 3 composes a text-to-image prompt that integrates prior descriptions holistically while emphasizing the current property, then uses DALL·E 3 to generate outputs (Choi et al., 29 Jul 2025).

Reuse is operationalized through three templates. Reuse Block places a saved block from the sidebar into a new path with original settings. Copy Path either replicates blocks and directions exactly in literal mode or uses GPT-4o to produce adapted directions consistent with new upstream context in context-adaptive mode. Recommend Next Directions produces two recommendations for a new block’s direction: one typical and one unique, based on exploration history and current canvas context (Choi et al., 29 Jul 2025).

The operator sketch is compact:

CreateBlock(p, d, r) → b
Suggest(b) → S
Generate(b, context) → O
Chain(b_prev, b_new)
CopyBlock(b) → b’
CopyPath(P, context, mode)
RecommendNext(property, history, context)

This procedure locates the model between prompt engineering and workflow orchestration. It complements systems that expose semantic axes or spreadsheet-based prompt construction, such as DreamSheets (Almeda et al., 2023), and systems that let users specify diversity bands and branching prompts, such as Expandora (Choi et al., 2 Mar 2025). Its novelty lies in binding diversified suggestions, context preservation, provenance, and reuse into a single block-and-graph formalism.

5. Evaluation and measured behavior

The model was evaluated in a within-subjects study with 12 designers against a baseline. Participants created significantly more input blocks with IdeaBlocks, with means of $d \in \Sigma^*$ 2 versus $d \in \Sigma^*$ 3 and $d \in \Sigma^*$ 4. They explored $d \in \Sigma^*$ 5 more images, with means of $d \in \Sigma^*$ 6 versus $d \in \Sigma^*$ 7 and $d \in \Sigma^*$ 8, corresponding to a $d \in \Sigma^*$ 9 increase (Choi et al., 29 Jul 2025).

Visual diversity was measured via CLIP embeddings of generated images using maximum pairwise cosine distance across a session:

$r \in \{1,2,3,4,5\}$ 0

and

$r \in \{1,2,3,4,5\}$ 1

On this measure, the system achieved higher diversity, with means of $r \in \{1,2,3,4,5\}$ 2 versus $r \in \{1,2,3,4,5\}$ 3 and $r \in \{1,2,3,4,5\}$ 4, a $r \in \{1,2,3,4,5\}$ 5 gain (Choi et al., 29 Jul 2025).

Exploration structure was analyzed with linkography. Link entropy was higher, with means of $r \in \{1,2,3,4,5\}$ 6 versus $r \in \{1,2,3,4,5\}$ 7 and $r \in \{1,2,3,4,5\}$ 8, indicating more diverse and interconnected idea development. Average link distance was larger, with means of $r \in \{1,2,3,4,5\}$ 9 versus $1 = \text{conventional}$ 0 and $1 = \text{conventional}$ 1, indicating more revisiting across longer temporal or conceptual gaps. The number of connected components was lower, with means of $1 = \text{conventional}$ 2 versus $1 = \text{conventional}$ 3 and $1 = \text{conventional}$ 4, indicating more cohesive structures rather than isolated clusters (Choi et al., 29 Jul 2025).

Qualitative findings align with these metrics. The typicality slider made range expression easier, and suggestions served as previews for clearer intent articulation. Reuse was frequent, at $1 = \text{conventional}$ 5 actions per session and $1 = \text{conventional}$ 6 total across sessions, especially for visual properties such as style and palette. Participants reported higher perceived quantity and diversity of outcomes, while perceived creativity remained similar because of limited fine-grained controllability in convergence (Choi et al., 29 Jul 2025).

These results place the model in a broader cluster of systems that improve exploratory behavior by restructuring representation and workflow rather than only improving generation quality. DreamSheets reported that experts built multidimensional sheet systems and generated an average of 7,925 unique images (Almeda et al., 2023). Inkspire improved exploration and inspiration in a within-subjects study (Lin et al., 30 Jan 2025). Expandora increased prompt diversity and the number of prompts tried within a fixed session (Choi et al., 2 Mar 2025). DynEx increased divergent and convergent thinking ratings in exploratory programming (Ma et al., 2024). Across these systems, the recurring empirical pattern is that explicit structures for branching, variation, and history alter exploration behavior measurably.

6. Positioning, implications, and limitations

Within design-exploration research, the model extends mood boards and gallery browsing systems such as Dream Lens, MetaMap, SemanticCollage, and DesignPrompt by reifying exploratory intent as modular, reusable objects linked into a provenance graph. It also complements prompt engineering tools such as Promptify, PromptCharm, and RePrompt by structuring intent and range, previewing suggestions, and chaining context-aware prompts to maintain continuity. Relative to provenance- and graph-based creative tools such as DeckFlow, Graphologue, XCreation, and GEM-NI, its stated novelty lies in modularizing Exploratory Intents into blocks with explicit typicality control, chaining that integrates prior visual context into subsequent prompts, and context-adaptive path reuse at scale (Choi et al., 29 Jul 2025).

The model also sits alongside a wider family of design-exploration formulations. DreamSheets frames exploration as sensemaking through local semantic and parametric axes in a spreadsheet interface (Almeda et al., 2023). Inkspire centers analogical sketching, per-stroke co-creation, and sketch-to-design-to-sketch feedback (Lin et al., 30 Jan 2025). Expandora formalizes prompt-branching with a novelty slider and a mindmap-like interface (Choi et al., 2 Mar 2025). StoryEnsemble represents the design process as a graph of persona, problem, solution, and storyboard nodes with forward and backward propagation (Suh et al., 5 Aug 2025). DynEx uses a Design Matrix to structure divergent and convergent reasoning before code generation (Ma et al., 2024). By contrast, optimization-oriented design space exploration in computer systems, HLS, CAD, or accelerator design is typically expressed through surrogate models, active learning, Bayesian optimization, or reasoning-driven simulator loops, as in HyperMapper 2.0, iDSE, AgentDSE, ForgeMorph, TRIM, and related work (Nardi et al., 2018, Li et al., 28 May 2025, Wang et al., 20 Jun 2026, Mazouz et al., 11 Apr 2025, Qi et al., 2021).

This contrast clarifies a second misconception: “design exploration” does not denote a single method class. In some literatures it denotes multi-objective search over formal design spaces; in others it denotes interactional support for divergence, continuity, and iterative reuse. The IdeaBlocks formulation belongs to the latter category, though it employs formal schemas, operators, and evaluation metrics.

The paper’s limitations are explicit. Sessions were 20-minute lab sessions, and participants were mostly design students or early-career designers. The domain scope was graphic and mascot design, so generalization to UI/UX, architecture, writing, music, or programming would require property schemas appropriate to each domain and potentially multimodal inputs. Convergence support remained limited, motivating future work on finer controllability, dynamic autonomy, and additional modalities such as sketching (Choi et al., 29 Jul 2025).

From these findings, several design implications follow directly. Early stages benefit from structured, system-driven breadth through typicality control and diversified suggestions. Intent should be represented as property, direction, and typicality, and supplemented with visual suggestions for abstract properties. Provenance-rich graphs and property-wise histories support revisiting, adapting, and branching without losing context. Process-aware recommendations that offer both typical and novel next steps can help sustain momentum (Choi et al., 29 Jul 2025). A plausible implication is that the Design-Exploration Model is best understood not as a fixed interface pattern, but as a reusable abstraction for making exploratory intents manipulable, persistent, and composable across non-linear creative workflows.