One Image is All You Need: Agentic One-Shot Image Generation via Text-Based World Models for Long-Tail Spatial Perception

The article introduces WMGen-v1, a novel agentic text-based world model framework designed for generating long-tail spatial data, crucial for applications like autonomous driving. This framework leverages a Large Vision-Language Model (LVLM) to create structured scene representations from a single reference image, while a Large Language Model (LLM) guides the scene expansion under physical and commonsense constraints. Benchmark results indicate that detectors trained on WMGen-v1 synthetic data can achieve performance comparable to those trained on real-world data, addressing the challenges posed by data scarcity in safety-critical scenarios.