Gen-AI-tecture

Two purpose-built generative-AI tools, developed for the first hours of the design studio

Generative AI Tools Curriculum Innovation Levels 4 & 5 Studio 2025–present >90% Creative Fluency

Overview

The early design studio has a familiar difficulty. The novice designer can model a room long before they can imagine one. They produce accurate geometry and empty atmosphere. Asked how the space should feel, they hesitate.

Two obstacles sit behind that hesitation:

The representational obstacle. A student who cannot yet draw fluently cannot yet test an idea quickly. The idea dies in the gap between conception and depiction.
The evaluative obstacle. A student who has seen few interiors has few references against which to judge their own. Precedent is the raw material of design judgement, and novices have little of it.

Gen-AI-tecture addresses both. It comprises two generative-AI applications that I designed and built for the studio, rather than adapted from commercial software. Each removes a specific obstacle at a specific moment in the design process. Neither produces the design.

The distinction matters, and it is the point of the whole endeavour. Gen-AI-tecture is not a standalone application. It is an expression of a commitment: that graduates should leave able to meet complex, real-world problems with innovative creative and technological approaches. The tools are the instrument. The graduate is the objective.

The Tools

1. Interior Style Transfer

The student uploads two images. The first is a view of their own three-dimensional CAD model, exported without materials. The second is a photograph of a real interior they consider relevant (Fig 1a–c, Fig 2a–c).

The tool transfers the visual language of the second onto the geometry of the first. Colour palette, materials, textures and light are carried across. The room itself is not redesigned. Walls, openings, joinery and furniture arrangement remain the student's own.

This constraint is deliberate. The student's authorship stays intact. What changes is only the thing they could not yet render for themselves.

Interior Style Transfer generative tool demonstration

Interactive Tool Demonstration · Interior Style Transfer Live video sequence showing student-authored untextured CAD geometry instantly mapped to the material, tonal, and light qualities of an uploaded reference interior photograph.

2. Interior Inpainting

The student uploads an image of an interior space. They then upload an image of a single piece of furniture. They paint a mask over the region of the interior where they wish the piece to sit (Fig 3a–d, Fig 4a–d).

The tool places the furniture within the masked region. Perspective, scale, shadow and reflection are matched to the host image. Everything outside the mask is left untouched.

The result is a fast, reversible test of a single decision. The student asks one question of their scheme, and receives one answer.

Interactive Tool Demonstration · Interior Inpainting Live canvas demonstration showing student-driven masking of a room region, contextual insertion of an isolated piece of furniture, and automated light and perspective synthesis.

Pedagogic Structure

Both tools were embedded in the Level 4 and Level 5 studios as staged exercises rather than as optional software. Four progressive stages structured their use:

Stage 1: Model before you generate. Neither tool functions without student input. The style-transfer tool requires the student's own CAD geometry. The inpainting tool requires the student's own interior and their own mask. No student can begin at the output.
Stage 2: Select the precedent, and defend it. Choosing the reference image is an act of design judgement, not a technical step. Students were required to state why a precedent was appropriate to their brief, their user and their site, before generating anything from it.
Stage 3: Generate, evaluate, discard. Outputs were treated as propositions. Students produced many and rejected most. The work of the studio moved from making an image to interrogating one.
Stage 4: Return to the model. The generated image is not the deliverable. Students were required to translate what the image revealed back into drawn and modelled decisions. Assessment rewarded that translation, and the reasoning behind it, rather than the visual quality of any generated output.

Learning Theory in Practice

The four stages were designed operationally. Read back, they express three theoretical positions:

Constructivism. Meaning is made by the learner, not transmitted to them. The tools give the novice a rapid cycle of proposition and appraisal at a stage when their own representational skill cannot yet sustain one. The student forms design judgement by exercising it, repeatedly and early.
Connectivism. Learning is understood as participation in a network, and that network now includes non-human participants. Students learned to work with a generative system: to frame a request, to read a response critically, and to know what the system could not tell them. This is a competence in its own right, and it is one they will exercise throughout their careers.
Universal learning. Design education has historically rewarded a narrow representational fluency. Students who draw confidently advance; students who do not are read as less able, when they are often only less practised. Both tools offer an alternative route from intention to image. Both function identically for the online learner and the studio-based one.

Outcome & Impact

Three findings emerged from use across the Level 4 and Level 5 cohorts:

Creative fluency was enhanced. Over 90% of students reported greater creative fluency at the early design stages when working with the tools than when working without them. The comparison is the significant part. Students were not asked whether they liked the software. They were asked to compare two ways of beginning a design.
Participation broadened. Two divisions were monitored: between students who described themselves as confident in drawing and those who did not, and between students attending online and those attending in person. All groups adopted the tools, and attainment gaps within the cohorts were minimal. Inclusion here is a measured outcome rather than an aspiration.
Confidence in AI-supported design was strengthened. Students left the studio having used generative AI as designers use any other instrument: deliberately, critically, and in service of a scheme they owned. They enter a profession in which such tools are becoming ordinary, already fluent and able to say what the tools are for.

Style transfer base model: student CAD kitchen layout — **Figure 1a · Base CAD Model**A view of a student-authored three-dimensional CAD model of a kitchen, exported without materials. Geometry, joinery, and layout remain strictly untouched.

Style transfer reference interior: tactile dark dining space — **Figure 1b · Reference Precedent**The reference image selected by the student: a dark, tactile dining space with textured plaster, live-edge timber, and warm low-level lighting.

Style transfer result: student kitchen rendered in precedent atmosphere — **Figure 1c · Generated Result**The kitchen returned in the atmospheric language of the precedent. Plan and cabinetry are unaltered; surface, colour, and light are transferred.

Style transfer base model: open-plan living and dining mezzanine loft — **Figure 2a · Base CAD Model (Loft)**A double-height loft space with a mezzanine stair and full-height glazed wall, exported without materials.

Style transfer reference interior: pale oak and linen — **Figure 2b · Reference Precedent**The chosen precedent: a pale, restrained interior of natural oak, linen, and soft diffused daylight.

Style transfer result: loft rendered in oak and linen palette — **Figure 2c · Generated Result**The loft rendered in the palette of the precedent. The comparison with Figure 1 demonstrates that the tool supplies no aesthetic of its own.

Inpainting base interior: brick arches living space — **Figure 3a · Base Interior**A student's living space with exposed brick arches and a boucle sofa, testing the addition of an occasional chair.

Inpainting furniture reference: tan leather armchair — **Figure 3b · Furniture Piece**The proposed piece: a tan leather armchair on a steel frame, supplied as a product photograph.

Inpainting mask: student painted region — **Figure 3c · Painted Mask**The mask painted by the student over the intended location, determining position, scale, and orientation.

Inpainting result: chair placed seamlessly in space — **Figure 3d · Inpainted Result**The chair placed within the masked region with matched perspective, scale, shadow, and reflection.

Inpainting base interior: bright timber ceiling living room — **Figure 4a · Base Interior**A bright living room with a folded timber ceiling and full-height glazing toward the garden.

Inpainting furniture reference: magenta buttoned chesterfield — **Figure 4b · Test Piece**A deliberately provocative test piece: a deep magenta buttoned leather chesterfield.

Inpainting mask: seating area replacement — **Figure 4c · Painted Mask**The masked region positioned along the glazed wall in place of the existing seating.

Inpainting result: chesterfield evaluated on sight — **Figure 4d · Inpainted Result**The chesterfield placed as specified, confronting the student with a tangible visual decision to assess and defend.

Significance

A novice designer's imagination should not be limited by their drawing hand

The approach rests on a single argument: a novice designer's imagination should not be limited by their drawing hand, and a studio can be built so that it is not. Four claims follow from the work.

Curriculum Innovation Founded on Development Rather Than Adoption

I did not select generative-AI software and fit teaching around it. I identified two specific obstacles in the early studio and built two tools to remove them. The pedagogic requirement preceded the technology, and determined its form.
Inclusion as a Measured Outcome

Adoption was consistent across learners of differing representational confidence, and across online and on-campus modes of study. Attainment gaps within cohorts were minimal. This is universal learning design demonstrated in practice, not asserted in a module handbook.
Futureproofing of Graduates

Students enter a profession being reshaped by artificial intelligence. They enter it having used such systems critically, within an assessed design process, and with a defensible account of what these tools contribute and what they cannot. That is an employability outcome of a durable kind.
Evidence-Based Guidance to the Discipline

The approach demonstrates how generative-AI workflows can be integrated into architectural pedagogy so as to operationalise constructivist principles of learner-led meaning-making, to support connectivist understandings of learning as participation in human–AI networks, and to advance universal learning theories through more inclusive, flexible and accessible educational practice.