README.md

metadata

license: mit
task_categories:
  - text-to-image
language:
  - en
configs:
  - config_name: default
    data_files:
      - split: combinations_common
        path:
          - combinations_layouts_common.json
      - split: combinations_uncommon
        path:
          - combinations_layouts_uncommon.json
      - split: count_2_4
        path:
          - count_layouts_2_4.json
      - split: count_5_7
        path:
          - count_layouts_5_7.json
      - split: count_8_10
        path:
          - count_layouts_8_10.json
      - split: position_boundary
        path:
          - position_layouts_boundary.json
      - split: position_center
        path:
          - position_layouts_center.json
      - split: size_large
        path:
          - size_layouts_large.json
      - split: size_small
        path:
          - size_layouts_small.json
pretty_name: LayoutBench-COCO

LayoutBench-COCO

Release of LayoutBench-COCO dataset from Diagnostic Benchmark and Iterative Inpainting for Layout-Guided Image Generation (CVPR 2024 Workshop)

[Project Page] [Paper]

Authors: Jaemin Cho, Linjie Li, Zhengyuan Yang, Zhe Gan, Lijuan Wang, Mohit Bansal

Summary

LayoutBench-COCO is a diagnostic benchmark that examines layout-guided image generation models on arbitrary, unseen layouts. Unlike LayoutBench, LayoutBench-COCO consists of OOD layouts of real objects and suports zero-shot evaluation. LayoutBench-COCO measures 4 skills (Number, Position, Size, Combination), whose objects are from MS COCO. The new 'combination’ split consists of layouts with two objects in different spatial relations, and the remaining three splits are similar to those of LayoutBench.

Skill Details

Skill 1: Number.

We define two layouts for 2∼10 objects and use 40 COCO objects, resulting in 720 total layouts (= 2×9×40). We name the layouts with 2∼4 and 8∼10 objects as few and many splits. The layouts are paired with captions with a template “a photo of [N] [objects]”.

Skill 2: Position.

For each of boundary and center splits, we define four layouts with 40 COCO objects, resulting in 320 total layouts (= 2 × 4 × 40). The layouts are paired with captions with a template “a photo of [N] [objects]”.

Skill 3: Size.

For each of tiny and large splits, we define nine layouts with 40 COCO objects, resulting in 720 total layouts (= 9×2×40). The layouts are paired with captions with a template “a photo of [N] [objects]”.

Skill 4: Combination.

This skill measures whether a model can generate two objects that commonly or uncommonly appear in the real world. For each of the three spatial relations (holding, next to, sitting on), we define three layouts without specifying objects. For each of the three relations, we manually define 20 object pairs of COCO objects for common and uncommon splits. For example, (1) ‘person sitting on chair’ is more common than (2) ‘elephant sitting on banana’ in real life. This results in 360 total layouts (= 2×3×3×20). The layouts are paired with captions with a template “[objA] [relation] [objB]”.

Use of LayoutBench-COCO

Step 1. Download LayoutBench-COCO and Generate Images from Layouts

Please see Step 1 README for instructions on downloading LayoutBench-COCO and generating images from layouts.

Step 2. Run Evaluation

Please see Step 2 README for instructions on running evaluation on LayoutBench-COCO.

Citation

@inproceedings{Cho2024LayoutBench,
  author    = {Jaemin Cho and Linjie Li and Zhengyuan Yang and Zhe Gan and Lijuan Wang and Mohit Bansal},
  title     = {Diagnostic Benchmark and Iterative Inpainting for Layout-Guided Image Generation},
  booktitle = {The First Workshop on the Evaluation of Generative Foundation Models},
  year      = {2024},
}