GPT-ImgEval: A Comprehensive Benchmark for Diagnosing GPT4o in Image Generation скачать в хорошем качестве

GPT-ImgEval: A Comprehensive Benchmark for Diagnosing GPT4o in Image Generation

This paper introduces *GPT-ImgEval, a new benchmark designed to thoroughly evaluate the image generation abilities of OpenAI's GPT-4o model* across three critical areas: general image creation, instruction-based image editing, and the synthesis of images requiring real-world knowledge. The researchers used existing and custom-built datasets (GenEval, Reason-Edit, and WISE) to quantitatively and qualitatively assess GPT-4o's performance, finding that it **significantly surpasses previous methods in both generating high-quality images with fine-grained control and understanding complex prompts**. Furthermore, the study delves into the potential underlying architecture of GPT-4o, presenting empirical evidence that it likely utilizes a **diffusion-based mechanism for image decoding**. The authors also identify several limitations of GPT-4o, compare its multi-round editing capabilities with Gemini 2.0 Flash, and reveal that its generated images, despite their realism, are **detectable by current AI forensic models**. Ultimately, GPT-ImgEval serves as a valuable resource for diagnosing GPT-4o's strengths and weaknesses and for guiding future research and development in the field of image generation. https://arxiv.org/pdf/2504.02782