New training method boosts diffusion model robustness against data contamination

Researchers have developed a new training method for diffusion models that enhances their robustness against data contamination. By replacing the standard Mean Squared Error (MSE) denoising loss with a transformation derived from f-divergences, the models show improved performance under corrupted datasets. This approach, termed divergence-induced weighted denoising, uses sample-specific influence weights to suppress errors from contaminated data, leading to better results on benchmarks like CIFAR-10. AI

IMPACT Enhances the reliability of generative models in real-world scenarios with imperfect data.

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

Read on arXiv stat.ML →

• CIFAR-10
• Diffusion Models
• f-divergence

• paper
• model release

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