A new study published in Fashion and Textiles proposes an alternative: instead of asking users to compose free-form prompts, a web-based system lets them pick design attributes from a structured taxonomy and hands the prompt-writing job to the machine itself. The research, led by Lidiia Khleborodova and Eung Tae Kim, frames the change as a shift from linguistic skill to variable selection, and reports preliminary evidence that the approach can make AI-assisted fashion ideation more transparent and more accessible to novices.
The problem the researchers set out to address is what the paper describes as a significant yet underexamined barrier in novice-oriented design settings. Generative AI systems built on GAN and diffusion models can visualize ideas rapidly, without advanced drawing ability, and prior studies suggest they can support divergent thinking during early design. But the dominant way of interacting with them, natural-language prompting, depends on what the authors call prompt literacy: the ability to choose appropriate vocabulary, manage attribute specificity, and anticipate how word choices map onto generated images. For users who are still developing a design vocabulary, the paper argues, this becomes a bottleneck that is structurally independent of their actual design knowledge. People may have ideas but lack the linguistic machinery to express them.
Template-based interfaces, which replace open prompts with predefined input fields, partially reduce that burden, but the authors contend they leave a deeper problem unresolved. In such systems, design attributes typically appear as independent slots with no explicit representation of their relationships, making it hard to see how individual decisions affect the output. The study therefore proposes a third paradigm it calls structured interaction, in which design attributes are parameterized as independent yet combinable variables, integrated through predefined compositional and constraint rules, and manipulated through controlled variation mechanisms. The defining feature is that relationships among attributes are built into the interaction architecture itself, so users can modify one variable while holding others constant and inspect the resulting differences.
The implemented system is a web-based fashion image generator built around six core attributes: item type, silhouette, color, material, pattern, and style. These were selected from fashion design literature, existing datasets and classification systems, and industry practice, then refined using professional references such as Fashionpedia and Stylepedia, with visual references drawn from runway archives. Each category was limited to roughly five to eight options, a cardinality constraint informed by prior findings that excessive representational complexity hinders processing in multi-variable tasks. Additional constraints ensured that options were visually and conceptually distinguishable, and that diversity remained balanced against control. The taxonomy and constraint design were reviewed by three experts with doctoral degrees in clothing and textiles, a process the authors describe as a check on clarity and interaction suitability rather than formal validation of the taxonomy as a classification system.
Under the hood, the system converts user selections into prompts through a rule-based translation pipeline. A constraint grammar defines the prompt as a fixed sequence of segments, including an item description, optional pattern and material descriptions, style conditioning, and standardized lighting, background, and framing elements. Deterministic mapping rules govern how each attribute becomes text: color, silhouette, and item type combine into a single phrase, style terms are expanded for semantic clarity, and pattern descriptions are restricted to the selected garment to prevent unintended propagation across the image. Because the prompt is assembled through a fixed, deterministic sequence, the system aims to minimize variation introduced by prompt formulation itself, allowing changes in user-selected attributes to be examined more directly. The authors distinguish this attribute-driven variation from stochastic variation, the randomness that appears when the same configuration is generated repeatedly.
Technically, the platform follows a client-server architecture, with a Vue.js frontend, a FastAPI backend, and image synthesis performed by Stable Diffusion models through the WebUI Forge API. To test how structured prompts fare across different diffusion architectures, the team ran a pilot exploratory evaluation using three model families: SD 1.5, SDXL, and Flux, represented by the checkpoints PerfectDeliberate v4.0, CyberRealistic XL v7.0, and Jib Mix Flux v12 SRPO. Three structured prompts were generated once per model, producing nine images in total. Semantic alignment was measured with attribute-level CLIP scores, visual quality with the LAION Aesthetic Predictor, and four graduate students in clothing and textiles rated each image on seven criteria using a seven-point scale. The authors are explicit that this was an indicative pilot, not a statistically generalizable validation of robustness.
Within those limits, Flux performed best. It achieved the highest CLIP score, 0.257, indicating stronger semantic alignment between structured prompts and generated images, and the highest LAION aesthetic score, 0.265. Expert ratings followed the same pattern: Flux received the highest mean scores on most fashion-specific criteria, including visual stability at 6.714, silhouette accuracy at 6.726, pattern correctness at 6.833, and overall photographic realism at 6.917, with relatively low variability across samples. Flux was accordingly chosen as the implementation model for the system. The team then ran an image generation analysis with the integrated model, examining sensitivity to single-attribute changes, stability across repeated generations, and behavior under exploratory combinations. Across 30 structured prompts evaluated at the element level, yielding 180 attribute assessments, 18 failures were observed, corresponding to an overall attribute reflection accuracy of roughly 90 percent. Most failures occurred in pattern, with eight cases, and material, with five, both of which demand fine-grained texture representation, while higher-level attributes such as color and style were consistently preserved.
The user study involved twenty undergraduate students from a department of clothing and textiles, all of whom had completed introductory fashion design coursework. Each participant completed an individual, guided session of roughly twenty minutes with the system, generating at least three design outputs and exploring variations by modifying individual attributes or regenerating identical configurations, followed by a questionnaire adapted from the Technology Acceptance Model, the System Usability Scale, and the Creativity Support Index. Ratings across four factors, perceived usefulness, ease of use, creativity support, and engagement, showed acceptable to excellent internal consistency, with Cronbach’s alpha between .79 and .89. A repeated-measures ANOVA found a significant main effect of factor type, F(3, 57) = 11.38, p < .001, partial eta-squared of .375. Post hoc comparisons with Bonferroni correction showed engagement rated significantly higher than usefulness and creativity support, and ease of use rated significantly higher than creativity support. Thematic analysis of open-ended responses echoed these results: participants praised the intuitive, step-by-step interface and the ability to iterate without writing prompts, but asked for broader option ranges, clearer visual previews, and better handling of subtle material and pattern features.
The authors are careful about what these findings do and do not show. The user study drew on a limited sample from a single institution, so results are framed as preliminary evidence of perceived usability and interaction interpretability, not generalizable conclusions about learning effectiveness or creativity enhancement. The absence of a direct comparison with free-form prompting means relative performance across interaction paradigms remains untested. The taxonomy itself is a simplified subset of fashion attributes, constructed for interaction purposes and not formally validated as a comprehensive classification system. The paper also notes a structural trade-off: repeated generation under identical attribute configurations tends to converge toward similar visual forms, so increased control comes at the cost of reduced diversity in design exploration. Even the environmental footprint of repeated image generation, the authors note, was outside the study’s scope but cannot be considered negligible.
Even with those caveats, the study offers a concrete demonstration that interaction design, not just model capability, shapes what generative AI can do in creative work. By embedding domain knowledge into the input representation, the framework makes the relationship between user input and generated output inspectable at the level of individual design decisions, rather than leaving users to reverse-engineer their own prompts. The authors suggest future work on hybrid interaction models that combine structured control with open-ended exploration, and on broader empirical validation across diverse user groups, tasks, and design contexts. For a field where the gap between having an idea and expressing it to a machine has become a genuine barrier, the message is that the interface itself can carry part of that expressive load.
Subject of Research: A taxonomy-driven interaction framework that replaces free-form prompting with attribute-based selection for AI-generated fashion images
Article Title: From prompting to attribute-based interaction: a taxonomy-driven framework for fashion image generation
Article References: Khleborodova, L., & Kim, E. T. (2026). From prompting to attribute-based interaction: a taxonomy-driven framework for fashion image generation. Fashion and Textiles, 13(1), Article 27. https://doi.org/10.1186/s40691-026-00477-8
Image Credits: AI Generated
DOI: 10.1186/s40691-026-00477-8
Keywords: generative AI, fashion design, diffusion models, human-AI interaction, prompt engineering, taxonomy, Stable Diffusion, user study, design ideation, attribute selection, usability, Flux
News Source: Denise Maddox. (October 30, 2026). Fashion AI tool swaps prompt writing for structured attribute selection. Scienmag.
