Research map / Vision-language and multimodal / Human and embodied modeling
APDDv2: Aesthetics of Paintings and Drawings Dataset with Artist Labeled Scores and Comments
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 APDDv2, an expanded painting and drawing aesthetics dataset with 10,023 images, expert attribute scores and language comments, plus ArtCLIP, a CLIP-based aesthetic scoring model fine-tuned on it.
Problem
Painting and drawing aesthetic assessment lacks large, richly annotated datasets; existing painting datasets have limited scoring dimensions and few annotations, and none include language comments.
Approach
Extends APDDv1 with more images, expert-labeled overall and 10 attribute scores under benchmark-based scoring standards, and detailed aesthetic comments; trains ArtCLIP with attribute-aware contrastive learning on comments (pretraining on categorized DPC2022) and fine-tunes an MLP regression head on APDDv2.
Key results
- ArtCLIP achieves total aesthetic score (TAS) SROCC 0.81, PLCC 0.84, ACC 0.89 on APDDv2, outperforming AANSPS (SROCC 0.76, PLCC 0.79, ACC 0.89) and SAAN (SROCC 0.78, PLCC 0.61, ACC 0.86).
- ArtCLIP obtains the best MSE and MAE on most attributes, e.g., TAS MSE 0.68 and MAE 0.63 versus AANSPS 0.88 and 0.73.
- APDDv2 expands APDDv1 from 4,985 to 10,023 images, with 85,191 scoring annotations and 6,249 language comments.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ArtCLIP instead of AANSPS | APDDv2 TAS evaluation · APDDv2 TAS | MSE: 0.88 -> 0.68 | Helped |
| Use ArtCLIP instead of AANSPS | APDDv2 TAS evaluation · APDDv2 TAS | MAE: 0.73 -> 0.63 | Helped |
| Use ArtCLIP instead of AANSPS | APDDv2 TAS evaluation · APDDv2 TAS | SROCC: 0.76 -> 0.81 | Helped |
| Use ArtCLIP instead of AANSPS | APDDv2 TAS evaluation · APDDv2 TAS | PLCC: 0.79 -> 0.84 | Helped |
| Use ArtCLIP instead of AANSPS | APDDv2 TAS evaluation · APDDv2 TAS | ACC: 0.89 -> 0.89 | No clear effect |
Surprising findings
- Due to the scarcity of painting commentary datasets, photography commentary datasets (DPC2022) were used for ArtCLIP pre-training instead.
Limitations
- Limited coverage of underrepresented or emerging art genres and cultural diversity of annotators.
- Only expert-level annotations; public aesthetic opinions not included.
- Painting commentary data scarce, so pretraining relied on photography comments (DPC2022).
Experimental setup
Datasets: APDDv2 DPC2022 APDDv1
Benchmarks: APDDv2
Metrics: MSE MAE SROCC PLCC ACC
Baselines: AANSPS SAAN
Builds on: CLIP APDDv1 AesCLIP AANSPS
Closest papers by meaning
- Bridging Cognitive Gap: Hierarchical Description Learning for Artistic Image Aesthetics Assessment (AAAI 2026 · similarity 0.88)
- AACP: Aesthetics Assessment of Children’s Paintings Based on Self-Supervised Learning (AAAI 2024 · similarity 0.87)
- ArtiMuse: Fine-Grained Image Aesthetics Assessment with Joint Scoring and Expert-Level Understanding (CVPR 2026 · similarity 0.86)
- Alchemist: Turning Public Text-to-Image Data into Generative Gold (NeurIPS 2025 · similarity 0.86)
- Stylebreeder: Exploring and Democratizing Artistic Styles through Text-to-Image Models (NeurIPS 2024 · similarity 0.86)
- VADB: A Large-Scale Video Aesthetic Database with Professional and Multi-Dimensional Annotations (NeurIPS 2025 · similarity 0.86)
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.