Datasets:
Facade — synthetic architectural style corpus
1,000 generated reference plates across 20 architectural styles, with structured attribute labels, generated style readings, and a precomputed retrieval index.
Built for a visual-similarity search task: photograph a building, retrieve the closest reference plates, get a reading of the style. Live app: Facade
What this is
Every plate is generated from a sampled attribute specification — style, massing, material, window rhythm, roofline, palette — plus controlled nuisance factors (viewing angle, crop, light, condition). Nothing is left to the generator's discretion.
That design buys three things:
- Ground truth for free. Every image inherits its labels from the specification, so retrieval evaluation needs no manual annotation.
- Measurable instruction compliance. Because each style declares an expected saturation band and dominant edge orientation, whether the generator obeyed is a computed quantity rather than an opinion.
- Sliceable nuisance factors. Angle, crop and light are controlled, so accuracy can be reported conditioned on them instead of averaged over them.
What this is not
The corpus depicts generic facades in a style. No plate represents a real building and none is attributable to any architect. Style labels describe the generation specification, not an art-historical judgement about any structure in the world.
Intended use is visual-similarity search. This is not an authoritative identification tool and carries no claim about any building's date, architect, or heritage status.
Files
| path | contents |
|---|---|
plates/ |
1,000 reference images, 512 px PNG, <style_id>-<nnn>.png |
plate_manifest.parquet |
per-plate labels, prompts, seeds and generated readings |
style_seed.csv |
the 20 style definitions and their expected measurements |
index_embeddings.npy |
precomputed image embeddings for the shipped model |
index_plate_ids.csv |
row order for the embedding matrix |
index_model.txt |
which model produced the index |
eda_summary.json |
measured compliance and signal statistics |
generation_log.csv |
per-attempt generation record (acceptance, rejection reasons) |
notebooks/ |
corpus generation, EDA, model comparison, publishing |
Loading
import numpy as np, pandas as pd
from huggingface_hub import hf_hub_download
REPO = "Jonathandav/facade-styles"
get = lambda f: hf_hub_download(REPO, f, repo_type="dataset")
E = np.load(get("index_embeddings.npy")) # (1000, d)
ids = pd.read_csv(get("index_plate_ids.csv")) # row order
manifest = pd.read_parquet(get("plate_manifest.parquet"))
styles = pd.read_csv(get("style_seed.csv"))
Plates resolve directly:
https://huggingface.co/datasets/Jonathandav/facade-styles/resolve/main/plates/ST01-000.png
The 20 styles
Bauhaus International · Brutalist · Art Deco · Ottoman Revival · Mandate Eclectic · Neoclassical · Gothic Revival · Venetian Gothic · Georgian · Victorian Terrace · Mid-century Modern · Postmodern · Deconstructivist · Contemporary Curtain Wall · Mediterranean Vernacular · Adobe Pueblo · Soviet Constructivist · Scandinavian Functionalist · Industrial Warehouse · Japanese Metabolist
Fifty plates each. Anything outside these twenty is still forced onto the nearest of them — worth knowing before trusting a result.
How it was built
| stage | detail |
|---|---|
| image model | SDXL-Turbo, 6 steps, guidance_scale=3.0, 512 px |
| text model | Qwen2.5-1.5B-Instruct, for the style readings |
| acceptance | every plate measured after generation; rejects regenerated with a fresh seed, up to 5 attempts |
Prompt construction is deliberate. The style name, period and iconic features
lead the prompt, because diffusion weights early tokens most heavily.
prompt_2 is not passed: in SDXL the second encoder produces the pooled
embedding that conditions the whole image, and putting a generic photographic
style there collapsed every architectural style into the same contemporary
office block.
Measured properties
Style compliance
Each style declares an expected saturation band and dominant edge orientation. Both are computed from pixels and compared against the declaration.
| test | result |
|---|---|
| declared saturation rank vs measured | Spearman ρ = 0.67, p ≈ 1.5 × 10⁻¹³², n = 1000 |
| saturation within calibrated band | see eda_summary.json |
| edge orientation as declared | see eda_summary.json |
The saturation bands are calibrated from the corpus, not set a priori. The first attempt used absolute thresholds chosen before any image existed and scored 8% compliance — because measured saturation ran 0.30–0.85 while the "medium" band had been guessed at 0.22–0.50. What the style table actually asserts is a ranking, not a value on a calibrated scale, so the headline test is rank agreement and the bands were rederived from corpus quantiles.
Does style dominate the nuisance factors?
Variance in each measured feature attributable to each factor (η²):
| feature | style | light | crop | view | condition |
|---|---|---|---|---|---|
| saturation | 0.63 | 0.12 | ~0.01 | ~0.00 | ~0.00 |
| brightness | 0.66 | 0.18 | ~0.01 | ~0.00 | ~0.00 |
| contrast | 0.43 | 0.19 | ~0.01 | ~0.00 | ~0.00 |
| orientation ratio | 0.56 | 0.04 | ~0.00 | ~0.00 | ~0.00 |
| edge density | 0.63 | 0.07 | 0.02 | ~0.01 | ~0.00 |
| angle entropy | 0.63 | ~0.01 | ~0.00 | ~0.01 | ~0.00 |
Style dominates every feature. Crop, view and condition sit at essentially zero — which validates the sampler: those axes are cycled rather than drawn independently, so they cannot correlate with style. If they did, an embedding model could score well by learning framing instead of architecture.
Light is the only meaningful contaminant, and only on the photometric features, which is physically unsurprising. Contrast has both the weakest style margin and the strongest light effect, making it the least trustworthy of the six.
Retrieval benchmark
Image-to-image retrieval over the corpus, style as the label, chance ≈ 0.049.
| model | Recall@1 | Recall@3 | MRR | lift |
|---|---|---|---|---|
| SigLIP-B/16 | 0.972 | 0.994 | 0.983 | 19.8× |
| LAION-CLIP-B/32 | 0.970 | 0.991 | 0.981 | 19.8× |
| CLIP-B/32 | 0.955 | 0.988 | 0.972 | 19.5× |
These are at ceiling and the comparison cannot resolve. Plates of the same style share nearly identical prompt strings, differing only in angle, crop, light and condition, so same-style plates are near-siblings. Every slice — by crop, view, light or condition — falls between 0.975 and 1.000. The differences between models are inside noise.
Note also that material and period are functionally determined by style
in this corpus (one value each per style), so they are not independent label
axes. ornament_level is.
Most confused pairs
| true | retrieved | rate |
|---|---|---|
| Georgian | Victorian Terrace | 0.16 |
| Victorian Terrace | Georgian | 0.08 |
| Mandate Eclectic | Ottoman Revival | 0.06 |
| Brutalist | Bauhaus International | 0.04 |
Georgian ↔ Victorian Terrace was predicted before running any embedding model, from centroid distance in measured saturation/orientation space. Two British brick terraces a century apart, and both the pixel metric and CLIP agree they are the hardest pair.
The remaining confusions have a structure the pixel metric could not see: Bauhaus International acts as an attractor for the modernist styles — Brutalist, Soviet Constructivist and Japanese Metabolist all leak toward it. That is a semantic effect, not a geometric one.
Sim-to-real
The corpus is synthetic; queries are not. Evaluated on 17 hand-labelled photographs of real buildings in Tel Aviv, Jaffa, Paris and the Greek islands, retrieving against the wholly synthetic index.
| model | real R@3 | synthetic R@3 | gap |
|---|---|---|---|
| LAION-CLIP-B/32 | 0.94 (16/17) | 0.991 | −0.05 |
| SigLIP-B/16 | 0.88 (15/17) | 0.994 | −0.11 |
| CLIP-B/32 | 0.76 (13/17) | 0.988 | −0.22 |
Per style (LAION): Bauhaus 1.00 · Ottoman Revival 1.00 · Mandate Eclectic 0.83 · Neoclassical 1.00 · Gothic Revival 1.00 · Mediterranean Vernacular 1.00.
The headline finding. The synthetic benchmark ranked SigLIP first by 0.002 on Recall@1. On real photographs LAION wins and has less than half the domain gap. Selecting on the saturated benchmark would have shipped the weaker model. A benchmark at ceiling does not merely fail to discriminate — it discriminates wrongly.
Mandate Eclectic scores lowest, and it was also the group flagged as lowest-confidence during hand-labelling. Labelling uncertainty surfacing in the results is a reassuring sign about both.
Failure cases are architecturally sensible rather than random: a Bauhaus building retrieved as Soviet Constructivist (contemporaneous interwar modernism, strip windows, white render), and a night photograph of a Mandate building retrieved as Mediterranean Vernacular (both cream render with shutters).
Generator failures found during construction
Four, documented in notebooks/02_facade_eda.ipynb. Three were silent — no
exception, no warning, and output that looked entirely plausible.
| # | failure | how it was caught | silent |
|---|---|---|---|
| 1 | the pooled embedding overrode every style | Bauhaus, Brutalist and Gothic all rendered as the same beige office block | yes |
| 2 | detail crops destroyed the class signal | style unrecoverable from a cropped window grid | yes |
| 3 | key_features never reached the image prompt |
code review after failure 1 | no |
| 4 | prompts truncated at 77 tokens | validator token-budget check | yes |
A fifth was a measurement bug rather than a generation one: the orientation
classifier could only emit three of the five declared labels, so every style
declaring curved or diagonal scored zero and looked like a total generator
failure. Fixing the metric raised orientation compliance from 60% to 83%.
The general lesson, and the reason this section exists: a synthetic corpus needs instrumented acceptance criteria. Looking reasonable is not evidence of a working generator, and a broken metric can invent a failure as easily as a broken generator can hide one.
Known limitations
- Style boundaries blur where visual grammars genuinely overlap; the confusion matrix above reports which pairs.
- Synthetic-to-synthetic retrieval saturates and cannot rank models reliably — see the sim-to-real section.
materialandperiodare determined bystyleand are not independent labels.- The real-photograph evaluation is small (n = 17) and unevenly distributed; per-style figures are reported alongside the mean for that reason.
- Generated plates are stylistic composites and may combine period details that would not co-occur on a real building.
- One style's declared edge orientation disagrees with measurement. The declaration is probably wrong, and it has been left visible rather than fitted to the data.
Ethical notes
No plate depicts a real building or is attributable to any architect. The accompanying app states, in the interface, that its output is a stylistic suggestion carrying no claim about a building's architect, date, or heritage status. Both the image and text prompts explicitly bar the models from naming real architects or buildings and from asserting heritage status.
Citation
@misc{facade_styles_2026,
title = {Facade: a synthetic architectural style corpus with
sim-to-real evaluation},
author = {Jonathan Dav},
year = {2026},
url = {https://huggingface.co/datasets/Jonathandav/facade-styles}
}
- Downloads last month
- 91