license: apache-2.0
dataset_info:
features:
- name: image
dtype: image
- name: elo
dtype: int64
- name: __index_level_0__
dtype: int64
- name: category
dtype: string
- name: subcategory
dtype: string
splits:
- name: train
num_bytes: 298637443.176
num_examples: 1999
download_size: 290047395
dataset_size: 298637443.176
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
tags:
- t2i
- preference
- ranking
- rl
- image
pretty_name: 2k Ranked Images
2k Ranked Images
This dataset contains roughly two thousand images ranked from most preferred to least preferred based on human feedback on pairwise comparisons (>25k responses). The generated images, which are a sample from the open-image-preferences-v1 dataset from the team @data-is-better-together, are rated purely based on aesthetic preference, disregarding the prompt used for generation. We provide the categories of the original dataset for easy filtering.
This is a new modality of preference data for us, different from our typical pairwise preferences, that we wanted to try out. If this type of data is useful of interesting to you, please ❤️ the dataset. If there is community interest and this dataset gets 30 ❤️s we will rank the full set of 17k images!
Examples
Some of the most preferred images
A lot of the most preferred images represent beautiful and serene landscapes and nature scenes.
Some of the least preferred images
The least preferres images are more diverse, but often contain scary, malformed, or just plain boring subjects.
Rank your own images through our API!