T2I Generation Evaluation

T2I: Text-to-Image

Overview

Text-to-image generation evaluation asks whether an image follows a prompt and how well it satisfies other needs, such as visual quality, spatial relationships, text rendering, factual accuracy, diversity, and safety. Researchers use controlled benchmarks, human ratings, detector- or program-based checks, and vision-language models that score images with questions or rubrics. Evaluation is expanding from broad alignment and aesthetics scores toward fine-grained tests of composition, cultural representation, specialized knowledge, and real-world use. A central challenge is that automated scores can miss subtle errors or reflect their own biases, while optimizing one measure—such as fidelity or safety—can harm another.

Latest papers

CardsList

184 papers20 papers in the last four weeks, down 13% on the four weeks before. 0.2% of all new papers.

  1. Multimodal Language Models as Text-to-Image Model Evaluators

    May 1, 2025Jiahui Chen, Candace Ross, Reyhane Askari-Hemmat +5LLM-as-a-JudgeBenchmark Design

  2. Conditional Vendi Score: Prompt-Aware Diversity Evaluation for Generative AI Models and LLMs

    Nov 5, 2024Mohammad Jalali, Azim Ospanov, Amin Gohari +1Generative AI EvaluationLanguage Model Generation Evaluation

  3. VerifyPrompt: Towards Text-to-Image Models Verification Behind Black-Box APIs

    Oct 30, 2024Ji Guo, Rui Zhang, Guomin Lu +4Adversarial Prompt GenerationT2I Generation Evaluation