README.md

metadata

license: openrail
task_categories:
  - text-to-image
language:
  - en
pretty_name: PaintSkills
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: count_val
        path:
          - count/val_images/*.png
      - split: spatial_val
        path:
          - spatial/val_images/*.png
      - split: object_val
        path:
          - object/val_images/*.png
      - split: count_train
        path:
          - count/train_images/*.png
      - split: object_train
        path:
          - object/train_images/*.png
      - split: spatial_train
        path:
          - spatial/train_images/*.png

PaintSkills

• ICCV 2023 Paper: DALL-Eval: Probing the Reasoning Skills and Social Biases of Text-to-Image Generation Models
• GitHub Repository: https://github.com/j-min/DallEval

Dataset Summary

PaintSkills is a compositional diagnostic dataset that evaluates three visual reasoning skills (object/count/spatial) of text-to-image generation models.

PaintSkills consist of a set of (text prompt, scene configuration, GT image) for each of three skills, which are collected in three steps:

1. We define scene configurations for each skill, in which the objects, attributes, and relations are uniformly distributed.
2. We generate text prompts by creating templates with objects, numbers, and spatial relations.
3. We generate images from the scene configurations using a 3D simulator.

Skills

1. Object Recognition (Object).

• Given a text describing a specific object class (e.g., an airplane), a model generates an image that contains the intended class of object.

2. Object Counting (Count).

• Given a text describing M objects of a specific class (e.g., 3 dogs), a model generates an image that contains M objects of that class.

3. Spatial Relation Understanding (Spatial).

• Given a text describing two objects having a specific spatial relation (e.g., one is right to another), a model generates an image including two objects with the relation.

Splits

PaintSkills have train and test splits.

The train split has 23,250/21,600/13,500 scenes for object/count/spatial skills.

The test split has 2,325/2,160/2,700 scenes for object/count/spatial skills.

Citation Information

@inproceedings{Cho2023DallEval,
  title     = {DALL-Eval: Probing the Reasoning Skills and Social Biases of Text-to-Image Generation Models},
  author    = {Jaemin Cho and Abhay Zala and Mohit Bansal},
  year      = {2023},
  booktitle = {ICCV},
}