qsun1 / gpt-4v-text2img-evaluation

Dataset for paper "Exploring GPT-4 Vision for Text-to-Image Synthesis Evaluation"

Text-to-Image Generation Dataset

Dataset and code for our paper: "Exploring GPT-4 Vision for Text-to-Image Synthesis Evaluation"

Updates

• (1/14/2024) Add 3 coco segmentations, see human_subset, object_subset, overlap_subset.
• (1/13/2024) Add 2 more text-to-image synthesis results: stable diffusion-XL turbo and Muse, see sdxl-turbo, openmuse, respectively.
• (12/8/2023) Add 50 samples generated by stable diffusion and GLIDE.

Introduction

• Environment python=3.9.18 torch=1.12.1+cu113
• directories coco is the ground truth image from COCO2014 validation dataset, glide is the image generated from GLIDE, sdv1-5 is the synthesized image from stable diffusion v1-5. Each subset has 50 samples generated from same text prompts.

How to get the caption and correponding image paths?

import json
with open('caption.json') as f:
    list_caption = json.load(f)
    captions = [item['caption'] for item in list_caption]
    img_pths = [item['image'] for item in list_caption]