Holistic Evaluation for Interleaved Text-and-Image Generation
Minqian Liu, Zhiyang Xu, Zihao Lin, Trevor Ashby, Joy Rimchala, Jiaxin Zhang, Lifu Huang
TL;DR
<3-5 sentence high-level summary> InterleavedBench and InterleavedEval address a critical gap in evaluating interleaved text-and-image generation by providing a diverse, instruction-rich benchmark and a strong, reference-free, multi-aspect evaluation metric based on GPT-4o. The dataset spans 10 real-world use cases with context-based and context-free subsets, totaling 815 instances, and emphasizes instruction-following and cross-modal coherence. Experimental results show pipeline approaches (LLMs plus image generators) outperform integrated multimodal models, while InterleavedEval achieves higher agreement with human judgments than prior metrics, highlighting both progress and remaining challenges, especially for image coherence. The work offers a practical framework and resources to guide future development in holistic interleaved generation and its evaluation, with implications for real-world multimodal content creation and storytelling.
Abstract
Interleaved text-and-image generation has been an intriguing research direction, where the models are required to generate both images and text pieces in an arbitrary order. Despite the emerging advancements in interleaved generation, the progress in its evaluation still significantly lags behind. Existing evaluation benchmarks do not support arbitrarily interleaved images and text for both inputs and outputs, and they only cover a limited number of domains and use cases. Also, current works predominantly use similarity-based metrics which fall short in assessing the quality in open-ended scenarios. To this end, we introduce InterleavedBench, the first benchmark carefully curated for the evaluation of interleaved text-and-image generation. InterleavedBench features a rich array of tasks to cover diverse real-world use cases. In addition, we present InterleavedEval, a strong reference-free metric powered by GPT-4o to deliver accurate and explainable evaluation. We carefully define five essential evaluation aspects for InterleavedEval, including text quality, perceptual quality, image coherence, text-image coherence, and helpfulness, to ensure a comprehensive and fine-grained assessment. Through extensive experiments and rigorous human evaluation, we show that our benchmark and metric can effectively evaluate the existing models with a strong correlation with human judgments surpassing previous reference-based metrics. We also provide substantial findings and insights to foster future research in interleaved generation and its evaluation.