Assignment 05
ARCH 507 — Generative AI: Thinking and Designing with Machine Intelligence
The most celebrated image models can seem to make almost anything. That breadth is not neutral. It comes from averaging. Models compress huge amounts of visual culture into statistical tendencies, so what is common and repeated becomes easy to make, while specific materials, places, aesthetics and cultural references get flattened, misrepresented or left out.
This assignment started from the opposite position. Each student built their own dataset and trained a LoRA on FLUX.1 [dev] to insert a new bias into the model: a material, a formal logic, a way of drawing, or a place the base model didn't hold with enough force or specificity. Then they made twelve images that work as a sequence, showing the world their intervention opens up.
The point wasn't only to prove a model could be trained. It was to ask why this dataset needed to exist. What did the model fail to show? What were they trying to recover, exaggerate or invent? And once the new data was in, what changed?
Each project page sums up what the student set out to test, what they built and what they found, often including what failed. The LoRAs are free to download and use.
General-purpose text-to-image models are trained to approximate a very large distribution of images, so they reproduce its densest regions most reliably. This study treats that averaging as a measurable bias. It asks whether a small, purposive dataset, attached to a frozen base model as a low-rank adapter, can insert a counter-bias: a material, a formal logic, a mode of representation or a place that the base model under-represents. Each project is an independent intervention on the same base model, so the results can be read side by side.
Before collecting any data, each student probed the base model with prompts describing their target condition and documented where it failed: generic outputs, wrong materials or scales, missing references. These failed prompts serve as the control condition. Every intervention is argued against them.
Datasets are small and curated by hand, between 10 and 41 images per adapter, with most between 15 and 35. Sources include the students' own photographs and 3D models, archival and published photography, film stills and drawings. Selection prioritises consistency in the target attribute, for example a correct brick scale or a single viewing condition. Images were cropped where needed to isolate it, as in the mullion and dune sets. Some students split one idea across several datasets so that separate adapters learn separable attributes, such as style vs. form, or stone texture vs. line drawing vs. façade photography.
Captions decide what the adapter binds to its trigger and what stays controllable from the prompt. Four strategies were used, and each project page reports which:
Where a trigger was used, it is usually a rare token (C_at_el, architamer, stoneybaloney) prefixed to every caption, so it doesn't collide with concepts the model already holds.
Students tuned adapter strength, trigger placement and prompt structure against base-model outputs. The working strength ranges each student reported, from 0.3 to 2.0, are listed with every LoRA. Each project concludes in a sequence of twelve images, suggested as four phases of three:
The set has to show a repeatable system with a distinct identity, not a one-off effect. A written statement places each intervention both technically and conceptually. Projects are assessed qualitatively on five criteria:
To isolate each adapter's effect, one prompt is rendered through the base model and all 25 adapters, with each adapter's trigger word added. Every other variable is held fixed:
A second pass sets each adapter to the middle of its student's recommended range. The results are under One prompt, every LoRA.
Some weights were submitted with machine-generated filenames. Each was identified by rendering same-seed tests with the student's own documented prompts and settings, then matching the results against their review material. Published files are renamed to clear names and checked by SHA-256 against the submitted originals.
Adapters are released on Hugging Face under the FLUX.1 [dev] Non-Commercial License, credited to each student. Training images are published in full only where they are the student's own work, in the public domain or openly licensed. Other datasets appear as captions, source lists and small thumbnails.

Teaching a razor-sharp model the grain, blur and quiet of old photographs.

A LoRA for traditional bamboo joinery: lashing, dowels and culm bundles across vernaculars.

Three LoRAs trained on film, old magazines and photography, to pull AI images back toward real design.

Does a culturally specific dataset remove a model's Western bias? Not entirely.

Using a flat, collage-like drawing style to escape the smudged look of "realistic" AI images.

A failed render LoRA, then a pencil-drawing LoRA that imagines the sustainable buildings of tomorrow.

Two LoRAs to get parametric brick facades with bricks at the right scale.

Can a LoRA rebuild a real, demolished building, not just something that looks like it?

Using dune geometry, not sand texture, to bend a facade's mullion grid.

Generating architectural elevations in the flat, colourful style of Sander Patelski.

Rwanda's volcanic stone, a material the base model reduces to a hut, carried into contemporary form.

Speculating on future plant species in Michigan's Upper Peninsula as the climate shifts.

Ancient and futuristic, sacred and technological: can AI hold opposites in one coherent scene?

Against the average: teaching the model the yellow stone, stairs and open doors of As-Salt, Jordan.

Style + form = ? A SANAA-style LoRA and a LoRA of Yewon's own product design, mixed to render a planter lamp.
One prompt, run through plain FLUX.1 [dev] and then through each of the 25 LoRAs, with the same seed and settings and each LoRA's trigger word added. It shows what each dataset pushes into the model on its own. Two versions: every LoRA at 1.0, and each at the strength its student recommended (the middle of their range).
an architecture magazine photo of an apartment inside the ruins of a park in central portland, designed in the style of peter zumthor and kengo kuma, with lots of concrete walls and cut-throughs to weathered stone structures. people are sitting at tables on an outdoor terrace overlooking the ocean.
FLUX.1 [dev] fp8 · seed 507507 · 28 steps · guidance 3.5 · euler / simple · 1344×768 · every LoRA at 1.0 · contact sheet (JPG)
![FLUX.1 [dev], no LoRA](assets/compare/strength-1/00_t.jpg)

























FLUX.1 [dev] fp8 · seed 507507 · 28 steps · guidance 3.5 · euler / simple · 1344×768 · each LoRA at its student's recommended strength (middle of their range) · contact sheet (JPG)
![FLUX.1 [dev], no LoRA](assets/compare/recommended/00_t.jpg)

























Not about AI replacing architects. Not about speed, optimisation or perfect results. Not about one correct way of working. Not about making more AI slop. Its ethics were part of the work: bias, plagiarism, agency and energy.
Guest reviewers: Angela Cho, Antje Steinmuller, Zain AbuSeir and Francesca Mavaracchio.
Every LoRA here was trained on FLUX.1 [dev] (rank 16, on fal.ai). None of them work with SDXL, even where a student's slides say SDXL.
Put the file in models/loras and load it with LoraLoaderModelOnly after the Flux model. Start at the strength the student used, listed on each project page.
Use the trigger word if there is one. Each page has an example prompt from the student's own tests.