Research map / Image generation / Text-to-image diffusion
Gender Artifacts from Art History to Text-to-Image Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces StyleGender, a ~74k-image dataset pairing WikiArt paintings and Stable Diffusion 3.5 / Flux generations across 19 styles with gender labels, plus two gender-artifact metrics (PixelSGA, MaskSGA) to measure how much gender is decodable from styles in art history versus T2I outputs.
Problem
Artistic style is treated in AI as a content-neutral visual filter, ignoring that historical styles encode gendered social roles; it is unclear whether T2I models carry these gender-coded patterns when prompted with style keywords.
Approach
Treats style as a carrier of gender artifacts: adapts the Meister et al. gender-artifact notion into set-level metrics (PixelSGA via downsampling/chromatic encodings and MaskSGA via subject/background occlusion) measured with a training-free kNN classifier on CLIP embeddings, and applies them to historical paintings and semantically aligned T2I generations.
Key results
- Neoclassicism and Art Nouveau score highest on art history (PixelSGA ~0.70/0.69, MaskSGA ~0.64/0.63); Symbolism is near chance (~0.52).
- All styles in SD and Flux generations show gender artifacts; PixelSGA and MaskSGA correlate moderately (Pearson r=0.51 SD, 0.64 Flux).
- On semantically aligned generations, SD and Flux show higher PixelSGA than historical counterparts in most styles, with Flux amplifying more consistently; MaskSGA shows exceptions (e.g., Surrealism, Pop Art).
Surprising findings
- MaskSGA is higher than art-historical values for generated images in several styles (e.g., Surrealism, Pop Art, Orientalism, Cubism, Contemporary Realism), contrary to the general amplification trend.
- Style keywords alone, without explicit gender semantics, produce systematic shifts in gender-related visual patterns.
Limitations
- SGA metrics depend on the kNN gender classifier's performance and inherit its limitations.
- Dataset has a Western bias (only Ukiyo-e is non-Western) and binary gender labels that may be ambiguous in abstract styles.
- Captions from Qwen2-VL may hallucinate or misidentify elements; some manual corrections were applied.
Experimental setup
Datasets: WikiArt StyleGender
Benchmarks: StyleGender
Metrics: PixelSGA MaskSGA balanced accuracy monotonicity score Spearman correlation
Baselines: SigLIP embedding variant (robustness check)
Builds on: Meister et al. gender artifacts DINO SigLIP Qwen2-VL
Closest papers by meaning
- Stylebreeder: Exploring and Democratizing Artistic Styles through Text-to-Image Models (NeurIPS 2024 · similarity 0.87)
- Beyond Aesthetics: Cultural Competence in Text-to-Image Models (NeurIPS 2024 · similarity 0.85)
- Alchemist: Turning Public Text-to-Image Data into Generative Gold (NeurIPS 2025 · similarity 0.85)
- Bridging Cognitive Gap: Hierarchical Description Learning for Artistic Image Aesthetics Assessment (AAAI 2026 · similarity 0.85)
- Rethinking Artistic Copyright Infringements In the Era Of Text-to-Image Generative Models (ICLR 2025 · similarity 0.84)
- Aligning Vision Models with Human Aesthetics in Retrieval: Benchmarks and Algorithms (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.