We introduce Text-to-Image Evaluation (T2IE), a state-of-the-art benchmark for evaluating text-to-image (T2I) models across eight key skills: counting, spatial relations, size composition, color accuracy, text generation, emotions, consistency, and typos. Our evaluation framework leverages a multimodal LLM (GPT-4o) to analyse generated images and assess their accuracy in relation to the corresponding skill. Additionally, for counting, size composition, spatial relations, text generation, and color accuracy, we integrate Detectron-2, a segmentation tool, to compare image outputs with prompts. To validate the benchmark’s reliability, 10 human evaluators conducted a manual assessment. We apply T2IE on three state-of-the-art T2I models: DALL-E 3, Adobe Firefly and Stable Cascade. Findings show that Stable Cascade excels in object counting but struggles with text generation and emotions. DALL-E 3 is consistent but has issues with spatial and size composition, yet it generates clearer text compared to Adobe Firefly and Stable Cascade. All the models show particular difficulties in text reproduction and emotional expressiveness, highlighting limitations in current T2I models and underscoring the need for advancements in both model development and evaluation methods.