EvalMuse-40K: A Reliable and Fine-Grained Benchmark with Comprehensive Human Annotations for Text-to-Image Generation Model Evaluation
Paper • 2412.18150 • Published
prompt_id stringlengths 5 5 | prompt stringlengths 3 401 | type stringclasses 2
values | img_path stringlengths 12 27 | total_score listlengths 3 6 | element_score stringlengths 179k 180k | promt_meaningless listlengths 3 6 | split_confidence listlengths 3 6 | attribute_confidence listlengths 3 6 | fidelity_label listlengths 0 23 |
|---|---|---|---|---|---|---|---|---|---|
00110 | A puffin sitting in booth while eating a pastry at a diner. Etching | real | SDXL-Turbo/00110.png | [
4,
3,
3
] | {"puffin (animal)":[1,0,1],"booth (object)":[1,1,1],"pastry (food)":[1,1,0],"diner (location)":[1,1,1],"sitting (activity)":[1,1,1],"eating (activity)":[0,0,0],"Etching (attribute)":[0,0,1],"pickle (food)":null,"mountain (location)":null,"standing (activity)":null,"top (spatial)":null,"scenic (attribute)":null,"realism... | [
0,
0,
0
] | [
0,
0,
1
] | [
1,
1,
0
] | [
"非真实场景",
"真实场景-动物"
] |
00110 | A puffin sitting in booth while eating a pastry at a diner. Etching | real | Playground_v2.5/00110.png | [
4,
3,
3
] | "{\"puffin (animal)\":[1,1,1],\"booth (object)\":[1,1,1],\"pastry (food)\":[1,0,1],\"diner (location(...TRUNCATED) | [
0,
0,
0
] | [
0,
0,
0
] | [
1,
1,
0
] | [
"非真实场景",
"真实场景-动物"
] |
00110 | A puffin sitting in booth while eating a pastry at a diner. Etching | real | Kandinsky3/00110.png | [
3,
3,
3
] | "{\"puffin (animal)\":[1,1,1],\"booth (object)\":[1,1,1],\"pastry (food)\":[1,1,0],\"diner (location(...TRUNCATED) | [
0,
0,
0
] | [
0,
0,
1
] | [
1,
0,
0
] | [
"非真实场景",
"真实场景-动物"
] |
00110 | A puffin sitting in booth while eating a pastry at a diner. Etching | real | IF/00110.png | [
3,
3,
3
] | "{\"puffin (animal)\":[1,1,0],\"booth (object)\":[1,1,0],\"pastry (food)\":[0,0,0],\"diner (location(...TRUNCATED) | [
0,
0,
0
] | [
0,
0,
1
] | [
1,
0,
1
] | [
"非真实场景",
"真实场景-动物"
] |
00110 | A puffin sitting in booth while eating a pastry at a diner. Etching | real | PixArt-Sigma/00110.png | [
3,
3,
3
] | "{\"puffin (animal)\":[1,1,1],\"booth (object)\":[1,1,1],\"pastry (food)\":[0,0,0],\"diner (location(...TRUNCATED) | [
0,
0,
0
] | [
0,
0,
0
] | [
1,
1,
0
] | [
"真实场景-物体",
"非真实场景",
"真实场景-动物"
] |
00110 | A puffin sitting in booth while eating a pastry at a diner. Etching | real | PixArt-alpha/00110.png | [
3,
3,
3
] | "{\"puffin (animal)\":[1,1,0],\"booth (object)\":[1,0,0],\"pastry (food)\":[0,1,0],\"diner (location(...TRUNCATED) | [
0,
0,
0
] | [
0,
0,
1
] | [
1,
1,
1
] | [
"真实场景-物体",
"非真实场景"
] |
00110 | A puffin sitting in booth while eating a pastry at a diner. Etching | real | SSD1B/00110.png | [
3,
3,
3
] | "{\"puffin (animal)\":[1,1,1],\"booth (object)\":[1,1,1],\"pastry (food)\":[0,0,1],\"diner (location(...TRUNCATED) | [
0,
0,
0
] | [
0,
0,
0
] | [
1,
0,
0
] | [
"真实场景-物体",
"非真实场景",
"真实场景-动物"
] |
00110 | A puffin sitting in booth while eating a pastry at a diner. Etching | real | SDXL/00110.png | [
3,
3,
3
] | "{\"puffin (animal)\":[1,1,0],\"booth (object)\":[1,1,0],\"pastry (food)\":[0,0,1],\"diner (location(...TRUNCATED) | [
0,
0,
0
] | [
0,
0,
1
] | [
0,
1,
1
] | [
"真实场景-物体",
"非真实场景",
"真实场景-动物"
] |
00110 | A puffin sitting in booth while eating a pastry at a diner. Etching | real | SD_v1.5/00110.png | [
2,
3,
3
] | "{\"puffin (animal)\":[1,1,1],\"booth (object)\":[0,0,1],\"pastry (food)\":[1,1,0],\"diner (location(...TRUNCATED) | [
0,
0,
0
] | [
0,
0,
1
] | [
0,
1,
1
] | [
"非真实场景",
"真实场景-动物"
] |
00110 | A puffin sitting in booth while eating a pastry at a diner. Etching | real | SD_v1.2/00110.png | [
3,
3,
3
] | "{\"puffin (animal)\":[1,0,0],\"booth (object)\":[1,1,0],\"pastry (food)\":[0,1,0],\"diner (location(...TRUNCATED) | [
0,
0,
0
] | [
1,
0,
1
] | [
0,
1,
1
] | [] |
End of preview. Expand in Data Studio
Dataset Card for EvalMuse
Dataset Structure
The repository includes images.zip.part-{aa-af}, which you need to merge manually. Use the following command to combine them:
cat images.zip.part-* > images.zip
unzip images.zip
- train_list.json: data used for training, containing detailed fine-grained annotations.
- test.json: data used for testing, annotations will be available later.
- prompt_t2i.json: selected 200 prompts and corresponding elements for image-text alignment evaluation of t2i models.
- split confidence , 1 means gpt based element split may got wrong result, while 0 stands for the gpt split the elements correctly.
- attribute confidence, 0 stands for the annotator was not sure about does the attribute keyword in the prompt match the corresponding image, since some attribute terms are hard to recognize for normal annotators.
- prompt meaningless, 1 stands for the prompt does not have major subject or the prompt is too abstract to evaluate text-image alignment.
Citation
@misc{han2024evalmuse40kreliablefinegrainedbenchmark,
title={EvalMuse-40K: A Reliable and Fine-Grained Benchmark with Comprehensive Human Annotations for Text-to-Image Generation Model Evaluation},
author={Shuhao Han and Haotian Fan and Jiachen Fu and Liang Li and Tao Li and Junhui Cui and Yunqiu Wang and Yang Tai and Jingwei Sun and Chunle Guo and Chongyi Li},
year={2024},
eprint={2412.18150},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2412.18150},
}
- Downloads last month
- 17
Repository:
github.com
Paper:
EvalMuse-40K: A Reliable and Fine-Grained Benchmark with Comprehensive Human Annotations for Text-to-Image Generation Model Evaluation
Number of rows:
32,717
Estimated number of rows:
32,717
Total file size:
54.6 GB