PromptMap: Map-Based AI Interaction

• PromptMap is a collection of map-based interaction paradigms that spatially organize prompt–image relationships to aid discovery.
• It supports exploratory design by visualizing precomputed prompt–image examples and non-linear prompt evolution using semantic maps and node-link histories.
• The approach enhances user exploration with intuitive subspace visualizations, facilitating systematic prompt refinement and improved creative outcomes.

PromptMap denotes a family of map-based interaction paradigms for generative AI, most prominently two 2025 systems for text-to-image generation that replace or supplement linear prompt entry with explicit spatial externalizations of the prompt space or the exploration process. One PromptMap system organizes a vast precomputed collection of prompt–image pairs into a semantically navigable 2D map with semantic zoom; the other models prompt exploration as a non-linear design process and visualizes prompts, images, and structured prompt subspaces as an exploration map. Both systems address a shared problem: users, especially novices, often struggle to articulate intent, discover viable prompts, and keep track of what has been tried (Adamkiewicz et al., 12 Mar 2025, Guo et al., 3 Oct 2025).

1. Scope and principal meanings

In current arXiv usage, PromptMap primarily refers to two related but distinct interaction styles for text-to-image systems. The first, introduced in "PromptMap: An Alternative Interaction Style for AI-Based Image Generation" (Adamkiewicz et al., 12 Mar 2025), treats the prompt space as a large, precomputed landscape of examples. The second, introduced in "PromptMap: Supporting Exploratory Text-to-Image Generation" (Guo et al., 3 Oct 2025), treats prompting as a design-exploration process and visualizes the evolving history of prompts, images, and structured variations.

Both systems are map-based, but the mapped entities differ. In the first system, the map contains example prompt–image pairs laid out so that nearby points are semantically similar, with density heatmaps, region labels, search, and semantic zoom. In the second, the map is a node-link representation of the user’s own exploration, with prompt nodes, image nodes, and subspace nodes that encode Cartesian products of prompt dimensions and values (Adamkiewicz et al., 12 Mar 2025, Guo et al., 3 Oct 2025).

Aspect	PromptMap (example-space)	PromptMap (process-space)
Primary unit	Precomputed prompt–image example	Prompt node, image node, or subspace node
Core structure	2D semantic map with zoom	Tree-like exploration map with nested grids
Main goal	Example-driven discovery and prompt adaptation	Tracking, structuring, and revisiting non-linear exploration

A plausible implication is that PromptMap is less a single interface than a design pattern: it externalizes otherwise implicit prompt relationships into spatial structures that users can navigate, compare, and reuse.

2. Semantic maps of prompt–image examples

The example-space PromptMap was proposed as an alternative to text-first prompting for novice users who struggle with vague intentions, "magic keywords," and search-plus-scrolling galleries such as DiffusionDB, lexica.art, and civitai (Adamkiewicz et al., 12 Mar 2025). Its central interaction is a map-like view over a vast synthetic corpus of examples. Each example prompt is a point; points with similar main subjects are placed close together; background shading is a density heatmap; and region labels annotate areas such as "space exploration" or "dragons and forests." Semantic zoom changes the representation by scale: zoomed-out views emphasize large density blobs and broad labels, intermediate views expose more fine-grained labels, and zoomed-in views fade the heatmap and reveal individual points and a random subset of thumbnails (Adamkiewicz et al., 12 Mar 2025).

The interface couples this map with a search panel, a prompt input panel for live generation, and a generated-images panel. Users can pan by dragging and zoom with the scroll wheel. Hovering a point reveals the prompt text, generated image, and attribute annotations such as likely location, subject, lighting, tone, mood, and genre; clicking pins the popup and supports copying the prompt to the clipboard. Search operates over annotation types including subject, location, lighting, mood, tone, and genre, and matching examples are highlighted as red points visible at all zoom levels (Adamkiewicz et al., 12 Mar 2025).

The intended workflow is example-driven prompt composition rather than prompt engineering from a blank text box. Observed usage included copying a prompt, pasting it into the live prompt box, and editing subject, style, or attributes; another recurring pattern was browsing multiple nearby examples to synthesize a new prompt from several inspirations. The system does not provide explicit multi-example composition, but it exposes prompt structure and attribute combinations directly enough that users can borrow and recombine them (Adamkiewicz et al., 12 Mar 2025).

This interaction style is explicitly exploratory rather than narrowly retrieval-oriented. The system is designed not only for finding an example close to an existing goal, but also for helping users "see what’s possible" by traversing neighborhoods of related visual ideas (Adamkiewicz et al., 12 Mar 2025).

3. Synthetic corpus construction, embedding, and indexing

A major technical contribution of the example-space PromptMap is its fully synthetic prompt corpus. Instead of scraping human-authored prompts, the system uses a recursive LLM-based pipeline centered on Mistral-7B-Instruct v0.2. GPT-4o first produces 160 general image categories. For each general category, the pipeline asks for 10 subcategories; for each subcategory, 10 sub-subcategories; and for each sub-subcategory, 20 idea captions. After deduplication, this yields about $247\,\text{k}$ unique idea captions across the 160 original categories and $11.6\,\text{k}$ unique subcategories. Each idea caption is then expanded into 10 location captions, each resulting pair into 5 main-subject captions, and finally into full prompts, producing approximately $12.35\,\text{M}$ prompts, reported as $12.4\,\text{M}$ in the paper (Adamkiewicz et al., 12 Mar 2025).

Deduplication is embedding-based. Intermediate outputs are embedded with CLIP; for each sample, the system finds 200 nearest neighbors with Faiss IVFPQ and removes neighbors with cosine similarity greater than $0.7$. After prompt synthesis, the LLM also predicts likely lighting, mood, tone, genre, and location tags for each prompt (Adamkiewicz et al., 12 Mar 2025).

The map layout is built from image-conditioned subject representations rather than prompt text alone. Each prompt is rendered with Stable Diffusion XL Turbo, and MiniCPM-V generates a main-subject caption for each image. These captions are embedded using all-mpnet-base-v2, then reduced to 2D with UMAP. The resulting coordinates are binned into a $2000 \times 2000$ histogram to form the density heatmap. Region labels are manually curated through anchor-point selection: an expert chooses label anchors, the system retrieves the 20 nearest points, and an LLM proposes a short label for that area (Adamkiewicz et al., 12 Mar 2025).

Search is implemented with Faiss IVFPQ indexes built separately for each annotation type. A query such as "flower-filled garden" is embedded with the same all-mpnet-base-v2 model, and the system retrieves 200 nearest neighbors from the chosen attribute index. Approximate ANN search enables real-time search over millions of examples on CPU. The broader stack uses Python 3 with Flask APIs, LMDB for fast random access to images, an NVIDIA A10 GPU for SDXL Turbo inference, CPU resources for search and map operations, roughly 20 GB RAM, and roughly 600 GB storage for images, embeddings, indexes, and metadata (Adamkiewicz et al., 12 Mar 2025).

The synthetic corpus was also evaluated as a dataset in its own right. Using the stable-diffusion-safety-checker, the paper reports 538 NSFW detections per 100k samples for the PromptMap synthetic dataset versus 4357 per 100k for DiffusionDB. In diversity experiments based on subject-caption deduplication, PromptMap’s recursive pipeline produced the highest number of unique subjects as sample size grew. Prompt length distributions also differed: PromptMap prompts had mean $M=17.2$, $SD=3.7$, whereas DiffusionDB prompts had mean $M=38.0$, $SD=22.3$ (Adamkiewicz et al., 12 Mar 2025).

4. Design-Exploration PromptMap and structured prompt subspaces

The second PromptMap system begins from a different diagnosis: users of text-to-image systems often become disoriented after generating many images because the exploration process is non-linear while standard interfaces present a linear history (Guo et al., 3 Oct 2025). The paper formalizes this with a Design-Exploration model built around experimentation, navigation, interpretation, and the structural notion of a subspace. A subspace is defined as a Cartesian product of dimensions and values. If $11.6\,\text{k}$0 is a set of dimensions and $11.6\,\text{k}$1 is the set of values for dimension $11.6\,\text{k}$2, then the subspace is $11.6\,\text{k}$3, with each combination corresponding to one prompt instance (Guo et al., 3 Oct 2025).

In this model, designs and subspaces interact dynamically. Users may begin with a single prompt or a structured subspace, experiment within it, elaborate it by adding dimensions or values, or branch to a new subspace when an alternative direction becomes promising. The paper explicitly contrasts this with data-frame theory: frames in sensemaking compete, whereas subspaces in exploration are complementary and are continually extended to cover more of the design space (Guo et al., 3 Oct 2025).

The interface represents this process as a node-link exploration map. Prompt nodes appear in three forms: an input form for editing a prompt and parameters; a prompt form that shows prompt text, parameter settings, and generated images; and an image form that isolates a single image dragged out from another node. Prompt differences relative to a parent node are highlighted. Subspace nodes use dimensional stacking as a nested grid representation: dimensions are assigned colors, the first two dimensions define a 2D grid, and additional dimensions recursively subdivide cells. Cells correspond to prompt configurations and can display generated images; their background color reflects aggregated user preference, with blue for liked and orange for disliked (Guo et al., 3 Oct 2025).

Interactions are centered on constructing and elaborating subspaces. A user can select part of a prompt, convert it into a dimension, and specify alternative values. The system then constructs a prompt template and instantiates all combinations of dimension values. Cells from a subspace grid can be dragged out to create new nodes with certain values fixed, after which further dimensions can be added. Nodes can be manually arranged on the canvas, pinned, minimized, or revisited through a mini-map. Image-level curation supports like and dislike operations, and liked images can seed new branches. The implementation uses Stable Diffusion SDXL as the backend text-to-image model (Guo et al., 3 Oct 2025).

The paper’s formative study identified three requirements that shaped this design: reflecting the evolution of user thinking, helping users compare and analyze the effects of prompt changes, and supporting efficient and structured exploration. The subspace abstraction addresses all three by making variations explicit rather than buried in a long prompt history (Guo et al., 3 Oct 2025).

5. Evaluation and empirical findings

The example-space PromptMap was evaluated with a between-subject online study with $11.6\,\text{k}$4 and a within-subject qualitative study with $11.6\,\text{k}$5. The quantitative study compared three conditions: No Support, Nearest Neighbor, and PromptMap. Post-task measures included NASA-TLX, Creativity Support Index, UMUX-Lite reported as SUS-parity, and three custom 7-point Likert questions on aesthetic quality, match to artistic vision, and satisfaction. The study found no statistically significant differences across the three interfaces on CSI, NASA-TLX, UMUX-Lite, the custom questions, or task completion times. Mean completion times nonetheless trended upward from No Support ($11.6\,\text{k}$6, $11.6\,\text{k}$7) to Nearest Neighbor ($11.6\,\text{k}$8, $11.6\,\text{k}$9) to PromptMap ($12.35\,\text{M}$0, $12.35\,\text{M}$1), which the paper interprets as more exploration rather than clear inefficiency (Adamkiewicz et al., 12 Mar 2025).

The qualitative findings for that system were stronger. Users shifted from trial-and-error prompt refinement in the baseline condition to example-driven exploration with Nearest Neighbor and especially PromptMap. Participants reported that examples helped them understand what was possible, exposed attributes they had not thought to specify, and reduced the sense of indirect, unpredictable control. Many preferred PromptMap’s synthetic examples to DiffusionDB examples because the synthetic prompts were perceived as longer, more detailed, and more diverse in topic coverage, and because thematic clusters plus search supported broader exploration beyond an initial query (Adamkiewicz et al., 12 Mar 2025).

The Design-Exploration PromptMap was evaluated through in-depth interviews with 8 frequent generative-AI users, including university professors in visual arts, a digital media student, amateur artists, and a professional information and interaction designer. Sessions were remote, lasted about 60 minutes, and used a Pair Analytics style in which participants controlled simpler interactions while the researcher assisted with more complex ones. Participants generally found the node-link plus nested-grid representation intuitive and legible. Reported benefits included clearer understanding of where ideas came from and how they developed, easier comparison of prompt variants, more systematic testing of synonyms and prompt dimensions, and substantially improved ability to revisit earlier branches and restore promising prompts (Guo et al., 3 Oct 2025).

The paper’s appendix also reports preliminary quantitative observations from session logs. On average, participants created about 5.89 prompt nodes between two subspace nodes, and subspace nodes constituted about 17.43% of all nodes. Two recurring patterns were identified: divergence-convergence cycles, in which users alternated between broad exploration and local refinement, and systematic fine-tuning, in which subspaces were used to compare small wording variations (Guo et al., 3 Oct 2025).

Across both PromptMap systems, the empirical record supports a consistent conclusion: map-based externalization changes workflow more clearly than it changes coarse usability metrics. This suggests that PromptMap’s main contribution lies in restructuring exploration and reflection rather than simply accelerating prompt entry (Adamkiewicz et al., 12 Mar 2025, Guo et al., 3 Oct 2025).

6. Relation to broader map-based and prompt-like research

PromptMap also sits within a broader lineage of map metaphors and structured spatial interfaces. "GMap: Drawing Graphs as Maps" (0907.2585) provides a direct precursor at the visualization level: it turns graph embeddings and clusters into geographic-like maps in which clusters become countries, the plane is partitioned with Voronoi-based regions, and users can interpret neighborhoods and higher-level groupings through a map metaphor. PromptMap applies a related intuition to prompt–image spaces, but with semantic zoom, example inspection, and live generation rather than static graph visualization [(0907.2585); (Adamkiewicz et al., 12 Mar 2025)].

In other arXiv literatures, the label "PromptMap" is used more loosely to denote systems that map complex inputs into structured, navigable representations. The MapQA paper explicitly describes a possible PromptMap system in which a user uploads a choropleth map and asks arbitrary natural-language questions. Its V-MODEQA architecture first extracts structured data from the map image and then applies table QA, and the authors present this as a blueprint for interactive map querying (Chang et al., 2022). PolyMapper is described as highly relevant to "PromptMap"-type ideas because it treats map extraction from overhead imagery as structured sequence prediction over ordered vertex tokens, directly predicting vector objects rather than dense segmentation masks (Li et al., 2018). MapQR extends the prompt analogy to autonomous-driving map construction: its scatter-and-gather query treats an instance query as a semantic prompt that is instantiated at multiple BEV reference points and gathered back into a coherent map-element representation (Liu et al., 2024).

These adjacent uses do not refer to the same HCI system, but they share a technical theme: complex spaces become more tractable when represented as explicit structures rather than opaque outputs. In one case the structure is a semantic landscape of examples; in another it is a graph of prompt history and subspaces; in others it is a table extracted from a map, a polygonal traversal of a road graph, or an instance-centric prompt to BEV features (Chang et al., 2022, Li et al., 2018, Liu et al., 2024).

Taken together, the PromptMap literature marks a shift away from strictly linear, text-box-centered, or pixelwise interaction styles. It favors representations in which users or models navigate organized spaces—semantic neighborhoods, subspace grids, graph-like histories, or prompt-like structured queries—and it treats those representations as first-class computational objects rather than as secondary visualizations (Adamkiewicz et al., 12 Mar 2025, Guo et al., 3 Oct 2025).