A tiny structural quirk in how AI handles empty space is the real reason it plagiarizes training images.
Image memorization in Stable Diffusion is driven by the way it handles padding tokens rather than the actual content of the prompt. When the AI sees empty space in a text command, it duplicates the end-of-text signal, which triggers a specific memorization pathway. This technical glitch, rather than a lack of training data, causes the model to spit out exact copies of its training set. Fixing this one specific token behavior could significantly reduce copyright issues in generative AI. It reveals that plagiarism in models is often a fixable architectural bug rather than an inherent flaw of the technology.
Memorization In Stable Diffusion Is Unexpectedly Driven by CLIP Embeddings
arXiv · 2605.02908
Understanding how textual embeddings contribute to memorization in text-to-image diffusion models is crucial for both interpretability and safety. This paper investigates an unexpected behavior of CLIP embeddings in Stable Diffusion, revealing that the model disproportionately relies on specific embeddings. We categorize input tokens as , , and with corresponding embeddings $\mathbf{v}^{\mathbf{sot}}, \mathbf{v}^{\mathbf{pr}}, \mathbf{v}^{\mathbf{eot}}, \mathbf{v}^{\mathbf{pad}}$. We discover th