New method delays diffusion model memorization, extends generalization

Researchers have developed a new method called Quality-Gated De-whitening (QGD) to control the memorization of training data in diffusion models. This technique aims to delay the point at which models begin to copy training examples, thereby extending the period of useful generalization. QGD works by maintaining a fast initial update phase and then gradually restoring momentum, with analysis separating different types of memorization. When tested on CIFAR-10 subsets, QGD significantly expanded the useful interval of diffusion models and reduced copying compared to standard SGD, while also achieving a better FID score. AI

IMPACT Improves the quality-copying tradeoff in diffusion models, potentially leading to more robust generalization.

RANK_REASON Academic paper detailing a new method for diffusion models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CIFAR-10
• Copy-Budgeted Selection (CBS)
• DagsHub
• Diffusion Models
• Gotit.pub
• Hugging Face
• IArxiv
• Influence Flower
• Quality-Gated De-whitening (QGD)
• ScienceCast
• SGD

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →