ViGoR-Bench: How Far Are Visual Generative Models From Zero-Shot Visual Reasoners?

1. 1Identify the limitations of current visual generative models in reasoning tasks
2. 2Develop a unified framework to evaluate visual generative models
3. 3Use ViGoR-Bench to assess the performance of models in zero-shot visual reasoning tasks
4. 4Analyze the results to inform future research and development directions

AI researchers and engineers working on visual generative models and computer vision tasks can benefit from ViGoR-Bench to identify areas for improvement, and product managers can use it to inform the development of more realistic benchmarks

💡 Current visual generative models struggle with tasks that require physical, causal, or complex spatial reasoning

🤖 ViGoR-Bench: a new benchmark to evaluate visual generative models' reasoning capabilities