GSI-Bench: First benchmark for generative spatial intelligence in multimodal models

GSI-Bench, introduced by Muzhi Zhu and team, is the first benchmark to quantify generative spatial intelligence (GSI) in multimodal models through spatially grounded image editing. It comprises GSI-Real (high-quality real-world data) and GSI-Syn (large-scale synthetic data with automated labeling). Fine-tuning unified models on GSI-Syn improves both synthetic and real spatial editing tasks, and also enhances downstream spatial understanding, showing generative training strengthens spatial reasoning.

• #multimodal
• #spatial-intelligence
• #benchmark
• #image-editing