---
title: "When Composition Doesn't Add Up: Humans Identifying Defects in AI-Generated Images"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.25933v1"
permalink: "https://feed.astralor.com/items/itm_8c5302303049d284690b6b59"
published: "2026-08-26T15:47:47.000Z"
score: 75
channel: "ai"
tags: ["ChatGPT","CO-AID","Future-IQA","Dataset","Text-to-Image Models","Compositional Defects"]
---

# When Composition Doesn't Add Up: Humans Identifying Defects in AI-Generated Images

## AI commentary

> This paper is worth attention as it provides the first large-scale manually annotated benchmark for compositional defects in AI-generated images, especially useful for engineers working on text-to-image model evaluation and tuning. Compared with manual spot checks or simple metrics, CO-AID offers a trainable foundation for defect localization and classification, potentially becoming a common evaluation resource in this field.

## At a glance

This paper addresses the pronounced and systematic defects of state-of-the-art text-to-image models when prompts involve intricate compositional factors such as multiple entities and attributes. The authors construct the CO-AID dataset by manually selecting 651 reference images from four categories (people, hand, object, scene) with complex composition, editing ChatGPT-generated prompts, feeding them to three T2I models, and having 29 participants provide multi-label defect annotations for each image. Experiments show that training a deep model on CO-AID can both predict defects and optimize AI image generation, demonstrating its usability and effectiveness.

## Key points

- Constructs the CO-AID dataset with reference images, prompts, AI-generated images, and defect location/type annotations.
- Covers four categories: people, hand, object, scene; each image annotated by 29 participants.
- Experiments show models trained on CO-AID can predict defects and improve generation quality.

## Full text

*Chulin Zhao and Ruoqi Hu contributed equally to this work. State-of-the-art text-to-image (T2I) models exhibit pronounced and systematic defects when prompts involve intricate compositional factors such as multiple entities and multiple attributes. In this paper, we investigate how humans identify such defects. Specifically, we manually select 651 reference images from the four categories of people, hand, object, and scene that exhibit complex compositional characteristics, from which prompts emphasizing compositional factors are derived by manually editing ChatGPT-generated prompts. We then feed the prompts into three selected T2I models to generate AI images and conduct a comprehensive subjective study to identify their defects. For each image, 29 participants provide multi-label assessments specifying defect types and locations. The study yields the compositional AI-generated image defect (CO-AID) dataset, including reference images, prompts, AI-generated images, and information on defect locations and types. Experimental results show that training a deep model on CO-AID can both predict defects in AI-generated images and optimize AI image generation, demonstrating its usability and effectiveness. The database and supplementary materials are available at: https://github.com/Future-IQA/CO-AID .

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_8c5302303049d284690b6b59