wenyii/PrefBench

We provide two prompt sources (diffusiondb and coco) and two metadata formats (json and pkl). Image files are stored in the corresponding diffusiondb/ and coco/ folders.

Path	Description
coco/	image files (e.g., coco/...png)
diffusiondb/	image files (e.g., diffusiondb/...png)
json/coco.json	One JSON object per user (COCO split)
json/diffusiondb.json	One JSON object per user (DiffusionDB split)
pkl/coco.pkl	PKL grouping used by PrefDisc-style trainers (see below)
pkl/diffusiondb.pkl	Same for DiffusionDB

Images

Extract image files with:

cat coco.tar.part-* | tar -xf - -C coco
cat diffusiondb.tar.part-* | tar -xf - -C diffusiondb

JSON

Each JSON file is a list of per-user records.
Each record is a dictionary with the following fields:

• image_file: path to the target image (training reconstruction target), e.g. coco/0000102215_0000066902.png.
• text: caption for the target image.
• negative_img: list of paths to dispreferred reference images.
• positive_img: list of paths to preferred reference images.
• prompt_list: list of prompts for the reference images.

For each record, negative_img, positive_img, and prompt_list are index-aligned and have the same length.

Example (abbreviated):

{
  "id": "0",
  "image_file": "diffusiondb/18869_0000001.png",
  "text": "pink, blue, despair personified, artwork",
  "negative_img": ["diffusiondb/18863_0036541.png", "..."],
  "positive_img": ["diffusiondb/18863_0000001.png", "..."],
  "prompt_list": ["...", "..."]
}

PKL

Each PKL file is a dict keyed by user.
Each value is a list that stores one user's reference pairs and attributes.

This PKL format contains the same preference information as JSON: reference pairs (negative_img, positive_img, prompt_list) plus negative_attributes and positive_attributes.

• For each reference pair: (negative_filename, positive_filename, prompt) (filenames can be joined with coco/ or diffusiondb/ under the dataset root).
• The last element of the list is [negative_attributes, positive_attributes].

In short, the value format is: [(neg_img1, pos_img1, prompt1), (neg_img2, pos_img2, prompt2), ..., [neg_attr, pos_attr]].

Example:

import pickle

with open("diffusiondb.pkl", "rb") as f:
    data = pickle.load(f)

sample_user = data[3]
print(sample_user)

Output:

[("145191_0004851.png", "145191_0000007.png", "a human chest burster coming out of a xenomorph"), ..., ["Academic Art ...", "Abstract Expressionism ..."]]