[Paper Review] Architecture inside the mirage: evaluating generative image models on architectural style, elements, and typologies

Jamie Magrill, Leah Gornstein+2 more|arXiv (Cornell University)|Jan 14, 2026

Aesthetic Perception and Analysis0 citations

TL;DR

The study evaluates five GenAI image platforms on 30 architectural prompts, measuring accuracy of generated images against historian-criteria, revealing limited overall accuracy and prompting labeling and provenance needs.

ABSTRACT

Generative artificial intelligence (GenAI) text-to-image systems are increasingly used to generate architectural imagery, yet their capacity to reproduce accurate images in a historically rule-bound field remains poorly characterized. We evaluated five widely used GenAI image platforms (Adobe Firefly, DALL-E 3, Google Imagen 3, Microsoft Image Generator, and Midjourney) using 30 architectural prompts spanning styles, typologies, and codified elements. Each prompt-generator pair produced four images (n = 600 images total). Two architectural historians independently scored each image for accuracy against predefined criteria, resolving disagreements by consensus. Set-level performance was summarized as zero to four accurate images per four-image set. Image output from Common prompts was 2.7-fold more accurate than from Rare prompts (p < 0.05). Across platforms, overall accuracy was limited (highest accuracy score 52 percent; lowest 32 percent; mean 42 percent). All-correct (4 out of 4) outcomes were similar across platforms. By contrast, all-incorrect (0 out of 4) outcomes varied substantially, with Imagen 3 exhibiting the fewest failures and Microsoft Image Generator exhibiting the highest number of failures. Qualitative review of the image dataset identified recurring patterns including over-embellishment, confusion between medieval styles and their later revivals, and misrepresentation of descriptive prompts (for example, egg-and-dart, banded column, pendentive). These findings support the need for visible labeling of GenAI synthetic content, provenance standards for future training datasets, and cautious educational use of GenAI architectural imagery.

Motivation & Objective

• Assess how well five widely used GenAI image platforms reproduce architectural styles, typologies, and elements from textual prompts.
• Quantify image accuracy using independent expert scoring across standardized criteria.
• Examine the impact of prompt frequency (Common vs Rare) on generated image accuracy.
• Characterize qualitative patterns in GenAI outputs to inform labeling and provenance standards.

Proposed method

• Use five GenAI platforms: Adobe Firefly, DALL-E 3, Google Imagen 3, Microsoft Image Generator, and Midjourney.
• Develop 30 architectural prompts spanning styles, typologies, and codified elements.
• Generate four images per prompt-platform pair (n = 600 images).
• Have two architectural historians independently score images for accuracy against predefined criteria; resolve disagreements by consensus.
• Summarize performance per set (0–4 accurate images per four-image set).
• Statistical comparison of Common vs Rare prompts (p < 0.05).

Experimental results

Research questions

• RQ1 What is the level of accuracy across GenAI platforms in reproducing architectural styles, typologies, and elements?
• RQ2 How does prompt frequency (Common vs Rare) influence output accuracy?
• RQ3 Are there platform-specific patterns in accuracy and failure rates?
• RQ4 What qualitative patterns emerge in GenAI architectural imagery that affect reliability and interpretability?

Key findings

• Mean accuracy across platforms is 42% (range 32%–52%).
• Common prompts yield 2.7x higher accuracy than Rare prompts (p < 0.05).
• Highest accuracy observed is 52%, lowest 32%; all-correct (4/4) outcomes are similar across platforms.
• All-incorrect (0/4) outcomes vary by platform, with Imagen 3 showing the fewest failures and Microsoft Image Generator the most.
• Qualitative patterns include over-embellishment, confusion between medieval styles and revivals, and misrepresentation of descriptive prompts (e.g., egg-and-dart, banded column, pendentive).
• Findings support visible labeling of synthetic content and provenance standards for training data; advise cautious use in education.