Alibaba T2I Review: Qwen Image Bench Open Source, GPT Image 2 Champion and Pentathlon According to Beating Monitoring, the Alibaba Qwen team announced the open source of a new drawing evaluation benchmark, Qwen Image Punch, specifically designed to evaluate the ability of large models to generate images from text (referred to as T2I, meaning automatic drawing of input text). A unified visual referee model Q-Judge based on Qwen3.6-27B deep training was also launched simultaneously. The evaluation benchmark simulates the professional art creation workflow, including five dimensions: image quality, aesthetics, text and image alignment, and newly added real-world fidelity and creative generation. It has 23 sub capabilities and 56 sub indicators. Qwen Image Bench contains 1000 bilingual stratified prompt words in both Chinese and English, with 500 long and 500 short descriptions, and an average of 4 or more dimensions assessed simultaneously. In order to conduct a detailed evaluation, the Q-Judge visual judgment model underwent blind review and triple review annotation under the supervision of 80 professional reviewers from art schools, with a training dataset covering over 130000 bilingual expert annotation pairs. The model outputs structured scores for 56 dimensions, with a high degree of agreement of 92% with human expert ratings. The evaluation results of the first batch of 18 mainstream image generation models showed that GPT Image 2 won the first place with a comprehensive score of 64.69 and ranked first in all 5 dimensions. Nano Banana 2.0 scored 59.82, GPT Image 1.5 scored 59.65, and Nano Banana Pro scored 59.45, ranking second, third, and fourth respectively. Alibaba's self-developed Qwen Image 2.0 Pro ranked fifth with 57.84, while GLM Image ranked last with 48.19. The data shows that real-world fidelity and creative generation are key indicators for opening up the model ladder. The evaluation also revealed the common technological bottlenecks in the current industry. AI painting models are generally prone to errors in drawing human bones, representing physical laws such as gravity and light and shadow, and handling details such as object interpenetration. Top models also score below 44 points in these dimensions. [Original link]