Memorization in Stable Diffusion is driven by the structural duplication of the CLIP <eot> embedding inside <pad> tokens, which causes over-reliance on that vector; simple inference-time masking or token replacement suppresses it without quality loss.
Finding dori: Memorization in text-to-image diffusion mod- els is not local
2 Pith papers cite this work. Polarity classification is still indexing.
2
Pith papers citing it
years
2026 2verdicts
UNVERDICTED 2representative citing papers
Diffusion models overfit denoising loss at intermediate noise but generalize in inference as model error smooths the flow field and sampling paths avoid memorized noisy training data.
citing papers explorer
-
Memorization In Stable Diffusion Is Unexpectedly Driven by CLIP Embeddings
Memorization in Stable Diffusion is driven by the structural duplication of the CLIP <eot> embedding inside <pad> tokens, which causes over-reliance on that vector; simple inference-time masking or token replacement suppresses it without quality loss.
-
Diffusion Models Memorize in Training -- and Generalize in Inference
Diffusion models overfit denoising loss at intermediate noise but generalize in inference as model error smooths the flow field and sampling paths avoid memorized noisy training data.