AI image generators build spatial relationships differently depending on how they learn language

Researchers found that diffusion transformers (the AI systems that generate images from text) use two completely different internal circuits to understand where objects should go in an image, depending on whether they learned language from scratch or from a pretrained model. The finding matters because one approach breaks more easily when given slightly different instructions — a signal that real-world image generation from text is fragile in ways the lab doesn't catch.

Image-generation AI works better in controlled settings than it does in messy reality. This paper shows why: the system's internal wiring for understanding spatial language is brittle. It can pass lab tests and still fail when users describe images slightly differently than the training data. The practical implication is that current models probably struggle with real user instructions more than benchmarks suggest — and that fixing it requires understanding these internal circuits, not just throwing more data at the problem.