New research indicates that standard accuracy metrics for diffusion models fail to detect critical numerical failures during the sampling process. This reveals a fundamental blind spot in how machine learning engineers validate the neural networks used in image generators like DALL-E and Stable Diffusion.