VisionFoundry: Synthetic Images Close VLM Visual Perception Gap in Spatial Understanding

arXiv·medium signalXBlueskyLinkedInCopy link

Zhou, Yin, and Chai show that vision-language models still struggle with visual perception tasks like spatial understanding and viewpoint recognition, and demonstrate that targeted synthetic image training data can close this gap. VisionFoundry generates task-specific synthetic images that teach VLMs perception skills they fail to learn from natural image distributions alone. Practical for teams fine-tuning VLMs for spatial reasoning, robotics vision, or document understanding applications.