---
configs:
- config_name: giaa
  default: true
  data_files:
  - split: train
    path: "giaa/train.csv"
  - split: validation
    path: "giaa/validation.csv"
  - split: test
    path: "giaa/test.csv"
- config_name: piaa
  data_files: "piaa/piaa.csv"
  keep_default_na: false
  na_values: [""]
- config_name: image_attributes
  data_files: "image_attributes/image_attributes.csv"
- config_name: image_prompts
  data_files: "image_attributes/image_prompts.csv"
pretty_name: LAPIS
license: cc-by-nc-4.0
---

# LAPIS: the Leuven Art Personalized Image Set
This repo is a cleaned version of the original LAPIS release -- see
[Data cleaning](#data-cleaning) below for exactly what changed and why.

<p align="center">
  <img src="img_labels.png" width="550" alt="image example with all its attributes">
</p>

LAPIS is the first dataset of artworks suitable for **personalized image
aesthetic assessment (PIAA)** i.e. predicting aesthetic ratings while
accounting for individual annotators' preferences, rather than a single
consensus score. It consists of 11,723 artwork images, meticulously
curated in collaboration with art historians, each with both an averaged
aesthetics score (GIAA) and the full set of individual ratings behind it
(PIAA), rich image-level attributes known to relate to aesthetic
appreciation, and rich personal attributes for each annotator.

- **Paper**: [LAPIS: a novel dataset for personalized image aesthetic assessment](https://arxiv.org/abs/2504.07670) (CVEU workshop, CVPR 2025)

## Contents

- **`images/`**: contains 2 folders, figurative and abstract, with the 11,723 images as plain `.jpg` files, split into figurative and abstract works according to the `2_styles` label. Download this folder either via the "Files and
  versions" tab above, `git clone`, or `huggingface_hub.snapshot_download`
  (see below).
- **`giaa/{train,validation,test}.csv`**: one row per image, aggregated
  across all raters (`min_response`, `max_response`, `mean_response`,
  `std_response`, `number of raters`, `mean_response_bin`). This is the
  "averaged ratings" table.
- **`piaa/piaa.csv`**: one row per individual rating (`rating`,
  `response time`, rater demographics, `participant_id`). This is the
  "individual ratings" table.
- **`image_attributes/image_attributes.csv`**: per-image low-level visual
  statistics (contrast, symmetry, fractal dimension, etc.), no rating data.
- **`image_attributes/image_prompts.csv`**: a generated caption (`prompt`) per image.

## Using the data without Python

Download `images/`, `giaa/*.csv` and/or `piaa/piaa.csv` directly from this
page (or `git clone` the whole repo). 
The `path` column (e.g.
`figurative/jan-matejko_....jpg`) is exactly where to find that row's image inside `images/`. The `image_filename` column is the image's bare filename, useful for matching a row across different tables.

**Note on `imgs_known`:** this column contains `None`, `1-10`, `11-25` or `more than 25`, and is empty for participants who didn't answer. Spreadsheet programs such as Excel may silently convert `1-10` and `11-25` into dates when you open the CSV directly. To avoid this, import the file via *Data → From Text/CSV* and set the `imgs_known` column to *Text*.

## Using the data for ML
```python
from datasets import load_dataset

giaa = load_dataset("asmaerten/LAPIS")                     # default config, all 3 splits
piaa = load_dataset("asmaerten/LAPIS", "piaa")
attrs = load_dataset("asmaerten/LAPIS", "image_attributes")
prompts = load_dataset("asmaerten/LAPIS", "image_prompts")
```

`load_dataset` keeps `imgs_known` values of `None` as text. If you read `piaa/piaa.csv` directly with pandas, pass `keep_default_na=False, na_values=[""]`. Otherwise pandas treats `None` as missing and merges it with the participants who didn't answer.

For the images, download the `images/` folder once and join by filename.
This works identically whether you're training on `giaa` (one row per
image) or `piaa` (one row per rating) -- just swap which table and which
rating column (`mean_response` vs. `rating`) you pass in:

```python
from pathlib import Path
from huggingface_hub import snapshot_download
from PIL import Image
from torchvision import transforms
import torch

images_dir = Path(snapshot_download("asmaerten/LAPIS", allow_patterns="images/*")) / "images"

class LapisDataset(torch.utils.data.Dataset):
    def __init__(self, table, images_dir, rating_column, transform):
        self.table = table
        self.images_dir = images_dir
        self.rating_column = rating_column
        self.transform = transform

    def __len__(self):
        return len(self.table)

    def __getitem__(self, i):
        row = self.table[i]
        image = Image.open(self.images_dir / row["path"]).convert("RGB")
        return self.transform(image), row[self.rating_column]

transform = transforms.Compose([transforms.Resize((224, 224)), transforms.ToTensor()])

giaa_train = load_dataset("asmaerten/LAPIS", split="train")
loader = torch.utils.data.DataLoader(
    LapisDataset(giaa_train, images_dir, "mean_response", transform),
    batch_size=32, shuffle=True,
)
```

## Data cleaning

This release differs from the original LAPIS release in a few ways:

- **Filename encoding fix**: 9 image filenames with accented names (Arnold Böcklin x6, Joaquín
  Sorolla x2, Grégoire Michonze x1) were handled inconsistently across different meta files, so a plain string join silently dropped these 9
  images. Normalized to a plain-ASCII slug consistent with how every other
  accented artist name in the dataset is already handled.
- **Participant 1075 removed**:
  This participant gave an identical rating for 20 consecutive trials. These trials should have been flagged and removed according to the exclusion criteria but were missed in the original release. After further inspection, we concluded this participant did not perform the rating task seriously throughout the majority of the experiment. We decided to remove all their ratings in this release. This affected 0.17% of the data.
- **40 images with duplicate style/genre labels resolved**: 40 images still had 2 different style or genre labels from the original wiki art meta data, which a
  merge step without an explicit join key fanned out into one row per
  candidate style/genre for every real trial on that image. Resolved by
  manual expert review, where the most suitable style/genre was chosen and the duplicate rows removed (2,205 rows, 0.78% of the data).

After these corrections, less than 1% of the data was removed. LAPIS now consists of 281,174 individual ratings, 11,723 images, averaging
23.98 ratings per image (range 13-67, median 24).
The initial release of the data can be found [here.](https://sites.google.com/view/lapisdataset/homepage)


## License

This dataset is licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/)
(non-commercial use, with attribution).

## Citation

```bibtex
@Inproceedings{Maerten2025LAPIS,
  title = {LAPIS: a novel dataset for personalized image aesthetic assessment},
  author = {Anne-Sofie Maerten and Li-Wei Chen and Stefanie De Winter and Christophe Bossens and Johan Wagemans},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
  year = {2025}
}
```