Gen-AI-tecture: using generative AI to support architectural students in design tasks

Synthesis: Kapsalis (2026) presents one of the first empirical studies of generative AI integration in architectural design education, using a locally executed, discipline-specific tool within a mixed-methods focus-group design. The study addresses three objectives: Creativity impact, inclusivity enhancement, and employability preparation. Results showed enhanced creative fluency, broadened participation across diverse learner profiles, and strengthened student confidence in AI-supported workflows.

This work extends the Generative AI education literature beyond text-based domains (coding, writing) into visual-spatial design disciplines. The finding that gen-AI tools broadened participation is particularly significant for Equity research — students who traditionally struggled with manual drafting or 3D modeling gained new entry points. This connects to AI Literacy discussions about AI as an Accessibility tool and Personalized Learning frameworks that emphasize multiple pathways to competence.

The emphasis on locally executed (non-cloud) AI is also noteworthy for Educational Measurement and Privacy-conscious deployment. The study operationalizes Constructivism principles by positioning AI as a tool for learner-led meaning-making within human-AI networks. For Educational Development, the paper provides evidence-based guidance on gen-AI integration in studio-based disciplines, an area where the The Evidence Base on AI in K-12: A 2026 Review and related literature have been thin.

Synthesis

The Gen-AI-tecture study shows that a bespoke, locally executed image-generation workflow can act as a constructionist "microworld" for architectural ideation rather than a mere image generator. Its convergent mixed-methods evidence — statistically strong correlations between questionnaire subscales and reflexive qualitative indices — indicates the tool most robustly expanded students' creative search space, operated as a modest equalizing mechanism for inclusive participation, and built immediate AI-handling confidence while leaving longer-term employability transfer more uncertain. The work demonstrates how a domain-specific, human-in-the-loop GenAI tool can operationalize Constructivism, UDL and connectivist principles within a studio Pedagogies and Teaching Strategies, extending largely text-focused AI-education research into the visual-spatial domain.

Key Findings

1. Creativity: students reported strongly enhanced idea generation (A1 = 4.5) and experimentation (A2 = 4.2), with the creative subscale correlating strongly with qualitative indices of divergent exploration (ρ = .76 for "more ideas than usual", ρ = .79 for "experimenting beyond usual repertoire"), though refinement of final decisions (A3 = 3.7) was less uniformly supported.
2. Inclusivity: the workflow operated as a modest equalizing mechanism — inclusivity items clustered high (3.8–4.1) and correlated strongly with feeling the session was "approachable/welcoming" (ρ = .74) and "keeping up regardless of prior AI use" (ρ = .81), with students less confident in drawing or with limited AI experience reporting they did not feel disadvantaged.
3. AI-handling and employability: in-session procedural confidence rose (C1 = 3.7, C2 = 3.5), but confidence in transferring these skills beyond the studio was markedly lower (C3 = 2.6), exposing a gap requiring sustained, curriculum-level provision.
4. Engagement: students produced roughly 80 images per session (~10 each) with acceptance rates above 60%, and used all three operation modes substantively — a modest preference for text-only prompting (Mode A) combined with continued reliance on reference-image and hybrid modes.

Background and Research Gap

The paper situates itself in a literature that has largely studied generic AI platforms rather than bespoke, discipline-specific tools aligned with architectural workflows and studio cultures. Prior work documents steep learning curves, fragmented toolchains and unpredictable outputs, pointing to the need for scaffolded training and explicit studio protocols. Survey evidence cited in the paper notes that nearly 70% of architecture students already use AI tools independently despite over 95% reporting no formal AI education, while the RIBA's 2025 "AI Report" finds over half of practices now integrate AI into projects — foregrounding employability stakes and the case for embedding AI Literacy into design education.

The Intervention: A Local, Discipline-Specific Workflow

The Gen-AI-tecture intervention centers on a bespoke image creation-and-editing workflow implemented as a localized ComfyUI node-based system, powered by the Flux 1 Kontext (dev) diffusion model fine-tuned via Low Rank Adaptation (LoRA) on the InteriorNet dataset of 20 million interior shots. The tool offers three modes — text-driven (A), reference-image-driven (B), and hybrid (C) — using in-painting to edit masked regions while preserving scale, perspective and context. Pedagogically it aligns with Constructivism principles and UDL by supporting learning-by-doing and multiple means of representation and expression, lowering technical barriers while still requiring reflection, judgment and iteration in line with studio pedagogy.

Study Design and Methods

The study adopts a focus-group, mixed-methods design with two groups (n = 8 each) of Level 3–5 architecture students at the University of Derby, purposively sampled for diversity in academic level, international status, declared disability, gender and prior digital-design familiarity. Sessions followed a stepwise 90-minute structure: warm-up and consent, a two-phase design task (Phase A create a master bedroom via Mode A; Phase B edit a living/dining room via Modes B/C), then a reflexive discussion and a post-session usability questionnaire. Data collection combined an eleven-item Likert "Usability Evaluation" questionnaire and semi-structured discussion, analyzed convergently — reflexive thematic analysis in NVivo (Braun & Clarke framework) plus descriptive statistics and Spearman's ρ rank-order correlations with Holm-adjusted p-values to test convergent validity between quantitative and qualitative indices. Around one-third of the sample declared a disability, including almost one-fifth reporting a specific learning difficulty, reflecting the project's inclusive focus.

Creativity Findings

The creative-skills subscale produced some of the highest scores. Students described expanded ideation — "normally I would stop after two or three options, but with the AI I felt I could quickly push five or six quite different room layouts before choosing one" — and stylistic experimentation across atmospheres and textures they would not usually draw. A3 (refinement) was more nuanced, with some students noting the AI helped them spot unconsidered combinations while others felt it could distract with details that did not fit the concept. The strong convergent correlations position the workflow as a cognitive artifact and "more knowledgeable other" that supports metacognitive talk about design moves and trade-offs.

Inclusivity Findings

Inclusivity-oriented items clustered in the upper-mid range, with students describing the tool as levelling: "Normally my rough sketches do not look like what is in my head, but here the AI helped me show the idea without needing perfect linework." The workflow reduced anxiety in crit-type situations and was experienced not as a bolt-on accommodation but as part of the "normal" design process — consistent with UDL's proactive, whole-cohort framing and with AI as an Assistive Technology for disabled and neurodivergent learners. The paper contrasts this with literature warning that AI can exacerbate inequities, arguing that a carefully scaffolded, locally controlled workflow can enable students to "keep up" irrespective of prior AI or drawing experience.

AI-Handling, Employability and Limitations

Students developed procedural confidence — "Seeing the mask, prompt and seed together helped me think of the AI as another tool in the workflow, not a magic black box" — but expressed uncertainty about transferring these skills professionally ("I am not yet sure I could explain the technical side confidently in an interview"). The study acknowledges limitations: a small, purposively sampled single-institution cohort; a single 90-minute exposure reliant on self-report; and a workflow fine-tuned for residential interiors that may not generalize to other design typologies or less supported environments. The paper calls for larger, multi-cohort, longitudinal and co-designed studies.

What this means for practice

• Instructors. Run GenAI studio sessions as a two-phase task — generate from a text prompt, then edit masked regions — so students rehearse prompting and iterative refinement rather than one-shot generation.
• Instructors. Keep the prompt, mask and seed visible while students work so the tool reads as an inspectable part of the design process rather than a black box.
• Designers. Build discipline-specific, locally run workflows around masked in-painting instead of general cloud text-to-image tools, so the scale, perspective and context of the design brief are preserved.
• Administrators. Treat a single workshop as an introduction only: the drop from in-session procedural confidence (C1 = 3.7) to transferable employability confidence (C3 = 2.6) signals that AI-handling competence for Workplace Learning needs curriculum-level provision.
• Researchers. Pair usability Likert items with reflexive discussion and report rank-order correlations to check that perceived creative gains match observed design behavior.

Limitations

• The mixed-methods focus-group design used a small, purposively sampled cohort of 16 Level 3–5 architecture students (two groups of 8) at a single UK institution (University of Derby), which the authors state constrains statistical power and transferability to other programs and disciplines.
• Evaluation captured a single 90-minute exposure and relied primarily on self-report questionnaires and reflexive group discussion, so the durability of perceived gains in creativity, inclusivity and AI-handling is unknown.
• No comparison condition was run: there was no generic-GenAI or non-AI control arm, so effects cannot be attributed to the bespoke workflow specifically.
• The workflow was fine-tuned for residential interior scenarios (Flux 1 Kontext LoRA on the 20-million-image InteriorNet dataset), and the authors state it may not generalize to other design typologies or less supported studio environments.

Citation

Timo Kapsalis (2026). Gen-AI-tecture: using generative AI to support architectural students in design tasks. Submitted to Journal of Architectural Education.