Memorization Detection in Diffusion Models via Text Embedding Interpolation

Text-to-image diffusion models exhibit memorization on certain prompts where the model reproduces a training image. A widely adopted detection approach quantifies the difference between the conditional and unconditional scores. We analyze this difference as a line integral of the conditional score along the linear interpolation from the unconditional embedding to the prompt embedding. Along this interpolation the score trajectory evolves smoothly over most of the interpolation segment and rises sharply within a narrow interval near the prompt embedding, producing the anomalously large score difference. We measure this sharp rise through the condition Jacobian evaluated at the prompt embedding, which achieves state-of-the-art detection accuracy and remains stable across reduced latent resolutions, enabling memory efficient detection by lowering the latent resolution at inference time.