Qwen/Qwen-Image-Bench · Hugging Face
Alibaba's Qwen team has released Qwen-Image-Bench, a vision-language model designed for evaluating text-to-image generated visuals. This model, fine-tuned from Qwen3.6-27B, assesses images based on a structured, hierarchical set of criteria including quality, aesthetics, alignment with prompts, real-world fidelity, and creative generation. Qwen-Image-Bench outputs its evaluations in a JSON format, utilizing chain-of-thought reasoning to provide detailed scores. AI
IMPACT Provides a new tool for automated assessment of text-to-image model outputs, potentially speeding up development cycles.