ARCH 507 · Winter 2026Taubman College · University of Michigan

Assignment 05

This Dataset Does Not Exist

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.

Methodology

1. Research question

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.

2. Baseline probing

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.

3. Dataset construction

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.

4. Annotation

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:

  • short hand-written captions;
  • long descriptive captions covering viewpoint, structure, material, light and objects;
  • captions drafted with vision-language models (Gemini, Claude, a Florence-2 workflow) and then edited by hand;
  • in one case, no captions at all.

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.

5. Training

  • Base model: FLUX.1 [dev], a 12-billion-parameter rectified-flow transformer.
  • Adapter: each LoRA learns a rank-16 update to the transformer's weights, and the text encoders stay frozen. At inference the update is scaled by a strength s, so W′ = W + s·ΔW.
  • Training run: adapters were trained with fal.ai's hosted FLUX LoRA trainer and exported as diffusers-format safetensors of about 86 MB each.
  • Output: 25 adapters from 12 students. Three further projects were submitted as images and write-ups without weights.

6. Inference and evaluation

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:

  1. establishing the inserted visual language;
  2. applying it across scales and conditions;
  3. transforming it through new material, formal or atmospheric possibilities;
  4. pushing it into speculative territory.

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:

  • the strength of the argument;
  • the specificity of the intervention;
  • the coherence of the sequence;
  • the fit between dataset, images and text;
  • the degree of authorship over generic generation.

7. Controlled cross-adapter comparison

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:

  • seed 507507;
  • 28 steps;
  • guidance 3.5;
  • euler sampler, simple scheduler;
  • 1344×768;
  • fp8 weights;
  • strength 1.0.

A second pass sets each adapter to the middle of its student's recommended range. The results are under One prompt, every LoRA.

8. Provenance and verification

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.

9. Release and data governance

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.

10. Limitations

  • Datasets are small and evaluation is qualitative. There is no held-out set or quantitative metric.
  • Trainer settings beyond rank were not recorded for every run.
  • Adapter strength works differently from one LoRA to the next. At a shared strength of 1.0 some adapters take over the image and others barely register, so cross-project comparisons should be read alongside each student's own tests.
Projects
15
LoRAs
25
Base model
FLUX.1 [dev]

Projects

Angelica Facey2 LoRAs

Memory

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

Arushi Chopra2 LoRAs

Bambu

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

Emily Naumann3 LoRAs

le_zine

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

Frank MichelImages + write-up

1970s Soviet Interiors

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

Hua Lin (Dylan)1 LoRA

2.5D World

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

Itamar Yahav2 LoRAs

pencil_drawing

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

Jack Barbour2 LoRAs

Future Facades

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

Jack SmithImages + write-up

Crockett · Villette · Oudolf

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

Leo Georges2 LoRAs

Sand

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

Michael Natinsky2 LoRAs

AI as a Drawing Tool

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

Ravi Kumar3 LoRAs

Amabuye

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

Riley Montgomery3 LoRAs

Upper Peninsula Understory Futures

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

Shuyang (Jesseco) ZhangImages + write-up

Liturgies of a Lost Future

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

Tamer Tubishat1 LoRA

Lithic Hospitality

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

Yewon Seo2 LoRAs

Render a Lamp

Style + form = ? A SANAA-style LoRA and a LoRA of Yewon's own product design, mixed to render a planter lamp.

One prompt, every LoRA

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
Base modelFLUX.1 [dev], no LoRA
Angelica Facey, C_at_el
Angelica FaceyC_at_el · 1
Angelica Facey, YO_c_me_T
Angelica FaceyYO_c_me_T · 1
Arushi Chopra, borsch_t
Arushi Chopraborsch_t · 1
Arushi Chopra, too_late
Arushi Chopratoo_late · 1
Emily Naumann, hoos
Emily Naumannhoos · 1
Emily Naumann, disasterpiece
Emily Naumanndisasterpiece · 1
Emily Naumann, minm
Emily Naumannminm · 1
Hua Lin (Dylan), Justdoit
Hua Lin (Dylan)Justdoit · 1
Itamar Yahav, pencil_drawing
Itamar Yahavpencil_drawing · 1
Itamar Yahav, peptides (first try)
Itamar Yahavpeptides (first try) · 1
Jack Barbour, Brick buildings
Jack BarbourBrick buildings · 1
Jack Barbour, Brick facade close-ups
Jack BarbourBrick facade close-ups · 1
Leo Georges, Mullions
Leo GeorgesMullions · 1
Leo Georges, Dunes
Leo GeorgesDunes · 1
Michael Natinsky, Patelski
Michael NatinskyPatelski · 1
Michael Natinsky, Elevations
Michael NatinskyElevations · 1
Ravi Kumar, stoneybaloney
Ravi Kumarstoneybaloney · 1
Ravi Kumar, katropalop
Ravi Kumarkatropalop · 1
Ravi Kumar, wikiwikiwoo
Ravi Kumarwikiwikiwoo · 1
Riley Montgomery, mont2
Riley Montgomerymont2 · 1
Riley Montgomery, lkjh
Riley Montgomerylkjh · 1
Riley Montgomery, mont3
Riley Montgomerymont3 · 1
Tamer Tubishat, architamer
Tamer Tubishatarchitamer · 1
Yewon Seo, asdf (style)
Yewon Seoasdf (style) · 1
Yewon Seo, qwer (form)
Yewon Seoqwer (form) · 1

About the course

Course aims

How the semester ran

  1. Building literacy: readings and discussion, with texts by Ranjodh Singh, Silvio Lorusso, William Morgan, Cleo Valentine and Cas Esbach.
  2. Testing the model: how diffusion models "sculpt noise", and the difference between a model and a product.
  3. Locating agency: ControlNet and other ways of steering the model, and using it as a tool versus as exploration.
  4. Inserting bias: the final project. Bias can't be removed, the real problem is hidden bias, and neutrality is false. So: what has the model normalised, removed, or never learned? Build the missing dataset, insert the bias, test the model.

What the course is not

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.

Final review

Guest reviewers: Angela Cho, Antje Steinmuller, Zain AbuSeir and Francesca Mavaracchio.

Using the LoRAs

Base model

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.

In ComfyUI

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.

Prompting

Use the trigger word if there is one. Each page has an example prompt from the student's own tests.