This study evaluates large language models' (LLMs) capabilities in comprehending political meme images across three dimensions: sentiment, rhetoric, and stance. A multidimensional evaluation framework was constructed, and expert annotations were performed on a X dataset. Results indicate that while LLMs show strong recognition, they still exhibit deficiencies in comprehending rhetoric and stance.In sentiment analysis, Gemini-2.5pro outperformed other models in emotional tendency classification (78.33% accuracy, F1 = 69.60%). However, all models showed low correlation in emotional intensity prediction, suggesting challenges with metaphorical ambiguity and cross-modal alignment. For rhetorical methods, Gemini-2.5pro led with a similarity of 0.4563, but all models performed below 0.46, revealing difficulty in capturing deeper semantics, especially for culturally sensitive devices like irony. In stance detection, Qwen2.5VL demonstrated unique strengths, leading with a similarity of 0.53165, while other prominent models like ChatGPT-4o and Gemini-2.5pro performed poorly.Overall, no single model consistently excelled across all tasks, and all task metrics had low upper limits. This highlights core bottlenecks in political meme parsing related to cultural context dependency, metaphorical polysemy, and cross-modal entity association deficiencies. This research offers theoretical foundations and practical insights for enhancing LLM application in multimodal political information analysis.