Research map / Image generation / Text-to-image diffusion
Synthetic History: Evaluating Visual Representations of the Past in Diffusion Models
Maria-Teresa De Rosa Palmini, Eva Cetinic
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DINOv2 ViT-B/14 linear probe instead of other encoders as style predictor (ablation) | style classifier · WikiArt 5-style validation | macro F1: 0.876 vs 0.868 (Swin-B), 0.658 (CLIP zero-shot) | Improved |
| rewrite prompt to ask for photorealism and add negative prompt 'black and white image' (ablation) | SDXL on HistVis · HistVis century-level styles | dominant style: SDXL still engraving-dominant in 17th-19th c. (17th c. VSD 0.93 -> 0.83) | No clear effect |
| rewrite prompt to ask for photorealism and add negative prompt 'black and white image' (ablation) | SDXL on HistVis · HistVis decade-level styles | monochrome proportion | Mixed |
| majority vote over three VLMs instead of a single VLM for anachronism detection (ablation) | anachronism detection pipeline · SD3 images user study | agreement with humans (%): 72 -> 75 (best single GPT-4o vs ensemble) | Improved |
| use GPT-4o rather than other LLMs for historical demographic baselines (ablation) | demographic baseline estimation · OWID comparison | MAE: 4.64 vs 4.83 (LLaMA-3.2), 6.93 (Claude 3.5), 7.66 (Qwen2.5 VL) | Improved |
| use open-source LLaMA-3.2 instead of GPT-4o to propose anachronisms (ablation) | anachronism proposal · HistVis | coverage | No clear effect |
| swap GPT-4o baselines for LLaMA-3.2 baselines in demographic analysis (ablation) | demographic alignment · HistVis | demographic deviation | No clear effect |
| cross-check FairFace with DeepFace ensemble | demographic extraction · 5,000 sampled images | agreement: 95.9% gender (kappa 0.93), 90.8% race (kappa 0.83) | No clear effect |
Closest papers by meaning
- Gender Artifacts from Art History to Text-to-Image Generation (NeurIPS 2026 · similarity 0.88)
- Beyond Aesthetics: Cultural Competence in Text-to-Image Models (NeurIPS 2024 · similarity 0.86)
- OASIS Uncovers: High-Quality T2I Models, Same Old Stereotypes (ICLR 2025 · similarity 0.86)
- The Hidden Language of Diffusion Models (ICLR 2024 · similarity 0.85)
- Diffusion PID: Interpreting Diffusion via Partial Information Decomposition (NeurIPS 2024 · similarity 0.84)
- Origin Identification for Text-Guided Image-to-Image Diffusion Models (ICML 2025 · 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.
