OpenAI’s latest GPT Image 2 model,It performed amazingly in the authoritative Wenshengtu evaluation, successfully surpassing Google Nano Banana2 and taking the first place in the world.This model was officially launched on April 21. After many months of iterative upgrades, the image quality, understanding and detail restoration capabilities have been greatly improved.
SuperCLUE evaluation data shows that,GPT Image 2 has reached the top in many core dimensions such as Chinese character generation, reality reproduction, and image quality, and its performance is very impressive.
Its Chinese character generation ability reaches 93.07 points, with full text accuracy. Whether it is seal script on blue and white porcelain or gold fonts on acrylic material, it can be perfectly integrated without a sense of suspension, completely solving the problem of garbled Chinese characters generated by overseas models.
Chinese character generation—image and text fit
Chinese character generation—image and text fit
In terms of scene restoration, the model can accurately reproduce complex scenes such as old-fashioned bakeries and intangible cultural heritage blacksmiths, with realistic and natural details. At the same time, he has a good understanding of long prompt words and logical reasoning requirements, and can accurately generate professional pictures such as scientific schematic diagrams and poster designs, and has a high degree of instruction compliance.
Physical scene reproduction
Compared with the previous generation product, GPT Image 2 has significantly improved image and text consistency and Chinese character generation, surpassing the previous generation model in all aspects. Compared with competing products such as Google and Baidu, it has also achieved leadership in multiple dimensions, especially in creative reasoning and reality reproduction.
At present, the model still needs to be optimized in aspects such as spatial relationship understanding and knowledge reasoning, but the overall model has reached the top level in the industry, marking a new stage of Vincentian graph technology.
Comparison of first-level dimension scores of domestic and foreign head models