FineGRAIN: Evaluating Failure Modes of Text-to-Image Models with Vision Language Model Judges

TL;DR

FineGRAIN presents a joint evaluation framework that probes both T2I and VLM systems through 27 failure modes using meticulously crafted prompts and LLM-guided judgment. The approach yields interpretable boolean and ranking scores and reveals nuanced failure patterns—especially in counting and text rendering—that existing metrics miss. A large dataset of multi-model T2I outputs, VLM annotations, and LLM judgments demonstrates that current evaluation metrics underrepresent failure risk, while FineGRAIN provides higher human-aligned assessment and actionable insights for reliability and interpretability in multimodal systems. This work offers a scalable, failure-mode-centric benchmarking paradigm to drive more robust, explainable generative vision-language models.

Abstract

Text-to-image (T2I) models are capable of generating visually impressive images, yet they often fail to accurately capture specific attributes in user prompts, such as the correct number of objects with the specified colors. The diversity of such errors underscores the need for a hierarchical evaluation framework that can compare prompt adherence abilities of different image generation models. Simultaneously, benchmarks of vision language models (VLMs) have not kept pace with the complexity of scenes that VLMs are used to annotate. In this work, we propose a structured methodology for jointly evaluating T2I models and VLMs by testing whether VLMs can identify 27 specific failure modes in the images generated by T2I models conditioned on challenging prompts. Our second contribution is a dataset of prompts and images generated by 5 T2I models (Flux, SD3-Medium, SD3-Large, SD3.5-Medium, SD3.5-Large) and the corresponding annotations from VLMs (Molmo, InternVL3, Pixtral) annotated by an LLM (Llama3) to test whether VLMs correctly identify the failure mode in a generated image. By analyzing failure modes on a curated set of prompts, we reveal systematic errors in attribute fidelity and object representation. Our findings suggest that current metrics are insufficient to capture these nuanced errors, highlighting the importance of targeted benchmarks for advancing generative model reliability and interpretability.

Figures (33)

• Figure 1: Overview of FineGRAIN architecture
• Figure 2: Comparison of agreement rates with human ground truth between VQAScore and FineGRAIN. FineGRAIN outputs a boolean prediction of whether the image contains a failure mode. VQAScore is a numerical score that we threshold to obtain a boolean (we ablate this threshold in the Appendix). FineGRAIN generally outperforms VQAScore.
• Figure 4: ROC curve depicting the performance of the LLM Judge. The curve illustrates the trade-off between true positive rate (TPR) and false positive rate (FPR) at various threshold settings.
• Figure 5: ROC curve for VQA Score evaluation, showing the classification effectiveness by plotting the true positive rate (TPR) against the false positive rate (FPR) for varying thresholds.
• Figure 6: Star plot comparing 5 diffusion modes on 27 failure modes as graded by a human evaluation