Comprehensive Performance Analysis: GLM-4.5V vs Gemini 2.5 Flash
Dataset: 128 EXPW Images | Ground Truth Comparison
| Model | Accuracy | Macro F1 | Weighted F1 | Precision | Recall | Tokens/Image | Speed Estimate |
|---|---|---|---|---|---|---|---|
| GLM-4.5V | 30.5% | 0.303 | 0.351 | 0.450 | 0.256 | 379 | 4.5s/img |
| Gemini 2.5 Flash | 67.2% | 0.569 | 0.681 | 0.581 | 0.574 | 1392 | 2.9s/img |
| GPT-5 Chat | 68.0% | 0.590 | 0.678 | 0.612 | 0.700 | 400 | ~3.5s/img |
GPT-5 Chat achieved the highest accuracy at 68.0%, making it the superior model for emotion recognition on this dataset.
Processing speeds vary across models, with Gemini 2.5 Flash being fastest (2.9s), GPT-5 estimated at ~3.5s, and GLM-4.5V at 4.5s per image.
GLM-4.5V uses approximately 73% fewer tokens per image (379 vs 1,392), significantly reducing API costs for large-scale processing.
Both models show varying performance across emotion categories. The macro F1 scores indicate room for improvement in detecting certain emotions like disgust and fear.