Dant33/WikiArt-81K-BLIP_2-captions

WikiArt Enhanced Dataset

Description

This dataset contains 81,444 artistic images from WikiArt, organized into different artistic genres. It has undergone several improvements and corrections to optimize its use in machine learning tasks and computational art analysis. Credits to the original author of daset go to: WikiArt

Enhancements

1. Encoding Issues Correction

• Fixed encoding issues in filenames and artist information.
• All filenames were renamed for consistency.
• Artist names were normalized, removing incorrect characters or encoding errors.

2. Genre Normalization

• Initially, some images belonged to multiple genres. The classification was reduced to a single main genre per image.
• Genres were consolidated into 27 main categories:

#	Genre	Size
1	Impressionism	13060
2	Realism	10733
3	Romanticism	7019
4	Expressionism	6736
5	Post Impressionism	6450
6	Symbolism	4528
7	Art Nouveau Modern	4334
8	Baroque	4240
9	Abstract Expressionism	2782
10	Northern Renaissance	2552
11	Naive Art Primitivism	2405
12	Cubism	2235
13	Rococo	2089
14	Color Field Painting	1615
15	Pop Art	1483
16	Early Renaissance	1391
17	High Renaissance	1343
18	Minimalism	1337
19	Mannerism Late Renaissance	1279
20	Ukiyo e	1167
21	Fauvism	934
22	Pointillism	513
23	Contemporary Realism	481
24	New Realism	314
25	Synthetic Cubism	216
26	Analytical Cubism	110
27	Action painting	98

• Genres were converted from list format (['Abstract Expressionism']) to string ("Abstract Expressionism").

3. Data Cleaning and Restructuring

• The description column was renamed to painting_name, removing hyphens for better readability.
• The columns width, height, and genre_count were removed.
• Missing values in phash, subset, genre, and artist were filled (2097 affected rows).
• phash values were generated for missing images.

4. Subset Reassignment

• The original distribution was:
  • Train: 63,998
  • Test: 16,000
  • Uncertain artist: 44
• A stratified 80/20 split was applied:
  • Train: 65,155
  • Test: 16,289

5. Automatic Description Generation

• A new description column was added, with descriptions generated using the BLIP2
• Example of the final dataset entry:

file_name,genre,artist,painting_name,phash,description,subset
Impressionism/william-merritt-chase_still-life-with-cockatoo.jpg,Impressionism,william merritt chase,still life with cockatoo,b0e24b85961e6de9,a painting of a white bird sitting on a vase,train

Final Dataset Format

The dataset is structured in a CSV file with the following columns:

• file_name: Image filename.
• genre: Assigned artistic genre.
• artist: Artist's name.
• painting_name: Painting title, cleaned and formatted.
• phash: Perceptual hash of the image.
• description: Automatically generated image description.
• subset: Subset to which the image belongs (train or test).

Usage

This dataset can be used for art classification tasks, automatic description generation, artistic style analysis, and training vision models.

License

This dataset is based on WikiArt and is intended for academic and research purposes.